Instructions to use sugiv/qwen3-8b-tanglish with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use sugiv/qwen3-8b-tanglish with Transformers:

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

pipe = pipeline("text-generation", model="sugiv/qwen3-8b-tanglish")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("sugiv/qwen3-8b-tanglish")
model = AutoModelForCausalLM.from_pretrained("sugiv/qwen3-8b-tanglish", 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 sugiv/qwen3-8b-tanglish with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "sugiv/qwen3-8b-tanglish"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sugiv/qwen3-8b-tanglish",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/sugiv/qwen3-8b-tanglish

SGLang

How to use sugiv/qwen3-8b-tanglish 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 "sugiv/qwen3-8b-tanglish" \
    --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": "sugiv/qwen3-8b-tanglish",
		"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 "sugiv/qwen3-8b-tanglish" \
        --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": "sugiv/qwen3-8b-tanglish",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use sugiv/qwen3-8b-tanglish with Docker Model Runner:
```
docker model run hf.co/sugiv/qwen3-8b-tanglish
```

qwen3-8b-tanglish / training /qwen_train_tanglish.yaml

sugiv

Add training config and full stdout log for reproducibility

0158250 verified 2 months ago

Raw

History Blame Contribute Delete

4.7 kB

	# Phase 4: Qwen3-8B LoRA fine-tune config (used by tanglish.llm.qwen_train).
	#
	# Strategy: full LoRA on bf16 base (NOT QLoRA/4-bit — L40S 48 GB has enough
	# VRAM for cleaner training with higher final quality).
	#
	# Never-lose-model design (per user 2026-07-15):
	# 1. All output goes to /workspace/tanglish/checkpoints/qwen_lora_v1/
	# which lives on the NETWORK VOLUME (survives pod eviction).
	# 2. A TrainerCallback S3-syncs every checkpoint AS IT'S SAVED
	# (not just at end).
	# 3. Post-training merge → S3 sync of merged model.
	# 4. Auto-terminate has 6 gates (train exited cleanly, no failed marker,
	# TRAINING COMPLETE marker, best-val LoRA on disk, merged model on disk,
	# S3 has both).
	# 5. Even mid-training pod eviction loses at most ~1 val-check-interval
	# of progress.

	# -----------------------------------------------------------------------------
	# Model
	# -----------------------------------------------------------------------------
	model:
	base: "Qwen/Qwen3-8B" # Qwen3 series has no "-Instruct" suffix; already instruction-tuned by default
	torch_dtype: bfloat16
	attn_implementation: flash_attention_2 # 2-3x faster than sdpa on Qwen3
	gradient_checkpointing: true # ~30 % VRAM save, ~15 % slower
	# Do NOT quantize — cleaner LoRA training on bf16 base; VRAM headroom exists.

	# -----------------------------------------------------------------------------
	# LoRA adapter
	# -----------------------------------------------------------------------------
	lora:
	r: 16
	alpha: 32 # 2x r is the standard heuristic
	dropout: 0.05
	bias: none
	target_modules:
	- q_proj
	- k_proj
	- v_proj
	- o_proj
	- gate_proj
	- up_proj
	- down_proj

	# -----------------------------------------------------------------------------
	# Data
	# -----------------------------------------------------------------------------
	data:
	# Downloaded from S3 in setup_qwen_train.sh, decompressed to plain .jsonl
	train_file: /workspace/tanglish/data/pairs/tanglish_v1.train.jsonl
	val_file: /workspace/tanglish/data/pairs/tanglish_v1.val.jsonl
	max_seq_length: 2048 # dialogues are short; leaves plenty of room
	packing: false # sample-level SFT (safer for chat format)

	# -----------------------------------------------------------------------------
	# Trainer — TRL SFTTrainer args (Hugging Face TrainingArguments compatible)
	# -----------------------------------------------------------------------------
	training:
	output_dir: /workspace/tanglish/checkpoints/qwen_lora_v1/lora

	num_train_epochs: 3
	per_device_train_batch_size: 4
	per_device_eval_batch_size: 8
	gradient_accumulation_steps: 4 # effective batch = 16

	learning_rate: 2.0e-4 # LoRA convention
	lr_scheduler_type: cosine
	warmup_ratio: 0.05 # 5 % warmup
	weight_decay: 0.01
	max_grad_norm: 1.0

	bf16: true
	optim: adamw_torch
	dataloader_num_workers: 4
	seed: 42

	logging_steps: 20 # wandb + local every 20 steps
	eval_strategy: steps
	eval_steps: 500 # ~10-15 min on L40S
	save_strategy: steps
	save_steps: 500 # save with eval so metric aligns
	save_total_limit: 3 # top-3 by eval_loss + last kept
	metric_for_best_model: eval_loss
	greater_is_better: false
	load_best_model_at_end: true # so merge picks the actual best

	report_to: wandb # step-level curves like Phase 2

	# -----------------------------------------------------------------------------
	# Early stopping — kills training when val loss plateaus
	# -----------------------------------------------------------------------------
	early_stopping:
	patience: 4 # 4 eval checks with no improvement
	threshold: 0.005 # must improve by ≥ this each time

	# -----------------------------------------------------------------------------
	# S3 sync callback — hits S3 after EVERY save event (paranoid but cheap)
	# -----------------------------------------------------------------------------
	s3_sync:
	enabled: true
	s3_prefix: "checkpoints/qwen3-8b-tanglish-v1/lora-ckpts/"
	# Also sync the merged model after training:
	merged_s3_prefix: "checkpoints/qwen3-8b-tanglish-v1/merged/"

	# -----------------------------------------------------------------------------
	# Wandb
	# -----------------------------------------------------------------------------
	wandb:
	project: qwen3-8b-tanglish
	run_name: v1 # override via CLI flag if you want