Qwen3-VL-Reranker-0.1B-tiny

This is a tiny version of Qwen/Qwen3-VL-Reranker-2B created for testing and development.

Intended Use

A small, fast stand-in for the qwen3_vl reranker architecture, useful for:

• Inference / CI testing where a real 2B checkpoint is too large to download or run
• Exercising the vLLM reranker -> Qwen3VLForSequenceClassification path via hf_overrides
• Quantization & compression pipeline smoke tests (llm-compressor, compressed-tensors)
• Offloaded / distributed loading tests (see below)

Weights are random (then briefly fine-tuned on a toy corpus), so scores/generations are not meaningful, this model is for plumbing, not output quality.

Model Details

• Base Model: Qwen/Qwen3-VL-Reranker-2B
• Architecture: qwen3_vl (Qwen3VLForConditionalGeneration), used in vLLM as the base for Qwen3VLForSequenceClassification
• Total Parameters: 0.099B
• Activated Parameters: 0.099B

Configuration Changes

The following parameters were reduced from the original model:

Attention head_dim is kept at 128, and the full 151,936-token vocabulary is retained.

Checkpoint Structure

Single safetensors file (model.safetensors). Key naming matches the original checkpoint format (model.language_model.*, model.visual.*). Module-path structure was verified equal to the base checkpoint's safetensors header.

Usage

Load as a sequence-classification reranker in vLLM (as with the full Qwen3-VL-Reranker-2B):

from vllm import LLM

llm = LLM(
    model="soyrsoyr/Qwen3-VL-Reranker-0.1B-tiny",
    hf_overrides={
        "architectures": ["Qwen3VLForSequenceClassification"],
        "classifier_from_token": ["no", "yes"],
        "is_original_qwen3_reranker": True,
    },
)

Or as a plain generative model in transformers:

from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained(
    "soyrsoyr/Qwen3-VL-Reranker-0.1B-tiny", device_map="auto"
)
processor = AutoProcessor.from_pretrained("soyrsoyr/Qwen3-VL-Reranker-0.1B-tiny")
input_ids = processor.tokenizer("According to all known laws", return_tensors="pt").input_ids.to(model.device)
print(processor.tokenizer.decode(model.generate(input_ids, max_new_tokens=20)[0]))

Offloaded / distributed loading (compressed-tensors)

This is a multimodal ...ForConditionalGeneration model, so AutoModelForCausalLM does not resolve it. Pass AutoModelForImageTextToText to load_offloaded_model — the class you pass must match the class you call, since that is where device_map="auto_offload" support is injected:

from transformers import AutoModelForImageTextToText
from compressed_tensors.offload import load_offloaded_model
from compressed_tensors.distributed import init_dist

init_dist()
with load_offloaded_model(model_class=AutoModelForImageTextToText):
    model = AutoModelForImageTextToText.from_pretrained(
        "soyrsoyr/Qwen3-VL-Reranker-0.1B-tiny",
        device_map="auto_offload",                 # weights on CPU/disk, GPU for activations
    )

Creation Process

This model was created using the llm-compressor create-tiny-model claude skill.

• Config inspected via inspect_config.py
• Tiny model created via a modified save_tiny_model.py, adapted for the multimodal class (AutoModelForImageTextToText.from_config); the text tower and vision tower were shrunk and any all-zero / non-finite / extreme param was fixed after init_weights()
• Fine-tuned on the copypasta dataset; reached training perplexity 1.00 (target: ≤3.0) at lr=5e-4 (CPU, Adafactor)
• Checkpoint structure validated against the original HuggingFace safetensors header (module-path match)
• Inference validated via validate_tiny_model.py

Notes

• Saved as Qwen3VLForConditionalGeneration (matching Qwen3-VL-Reranker-2B). vLLM converts it to Qwen3VLForSequenceClassification at load time via the hf_overrides shown above, so this tiny model exercises the reranker → sequence-classification path.
• Projector alignment. vision_config.out_hidden_size is set to the text hidden size (512) so the visual merger projects into the text tower; deepstack_visual_indexes is remapped to valid indices for the reduced 4-layer vision tower.
• tie_word_embeddings=True: lm_head shares embed_tokens and is not stored as a separate tensor.

Validation output: Success: 1.003219485282898 <= 10.0