Clef-Flash-W4A16-AutoRound-GPTQ

Quantized W4A16 (4-bit weights, 16-bit activations) release of Cloudflare/clef-flash using Intel AutoRound.

Clef-Flash is a 9B multimodal decision model post-trained from Qwen3.5 that evaluates structured schemas (text, JSON, image, video) and returns calibrated probability distributions across typed questions in a single forward pass without autoregressive token generation.

---

Quantization Details

• Method: Intel AutoRound W4A16 (sym=True, group_size=32)
• Vision Tower: Preserved in native BF16 (ensures zero OCR/visual degradation)
• Joint Schema Head (joint_head.safetensors): Preserved in native BF16
• Linear Convolutions: Preserved in native BF16 to prevent drift in Qwen3.5 linear attention layers
• Accuracy: Zero loss observed across text, JSON state, and multimodal test suites compared to the base checkpoint.

---

Quickstart

vLLM Serving (/v1/systemone)

Requires vllm > v0.31.0 or vllm/vllm-openai:nightly.

1. Launch Server

docker run --gpus all -p 8000:8000 --ipc=host vllm/vllm-openai:nightly \
  vllm serve Vishva007/clef-flash-W4A16-AutoRound \
    --port 8000 --host 0.0.0.0 \
    --max-model-len 32k --max-num-batched-tokens 32k --max-num-seqs 128 \
    --kv-cache-dtype fp8 --gpu-memory-utilization 0.90 \
    --enable-prefix-caching --enable-chunked-prefill \
    --trust-remote-code --dtype bfloat16

2. Client Example (typesafe_sdk)

pip install typesafe_sdk

from typesafe_sdk import Choice, TypeSafeClient

with TypeSafeClient(
    api_key="EMPTY",
    base_url="http://localhost:8000/v1",  # or your proxy URL
    model="Vishva007/clef-flash-W4A16-AutoRound",
) as client:
    res = client.system_one(
        state="Checkout throwing 504 Gateway Timeouts. Database latency at 4,000ms.",
        questions={
            "severity": Choice(
                instructions="Assess severity",
                criteria={"sev1": "Total outage", "sev2": "Degraded", "sev3": "Minor"},
            ),
            "team": Choice(
                instructions="Route to team",
                criteria={"sre": "Infrastructure", "billing": "Payments"},
            ),
        },
    )

    for k, v in res.choices.items():
        print(f"{k}: {v.choice} ({v.confidence:.1%})")

Installation

pip install torch transformers huggingface_hub pillow

Usage (systemone API)

import sys
from huggingface_hub import snapshot_download

# Download repo and load custom joint schema model
repo_id = "Vishva007/clef-W4A16-AutoRound"  # or auto-gptq / llm-compressor variant
model_path = snapshot_download(repo_id)
sys.path.insert(0, model_path)

from joint_schema_model import load_release_model, systemone

model, processor = load_release_model(model_path, device="cuda")

# 1. Text / JSON Decision Example
response = systemone(model, processor, {
    "model": "clef-flash",
    "state": "Prod database latency spiked to 4,000ms. Checkout failing with 504 Gateway Timeouts.",
    "questions": {
        "severity": {
            "type": "choice",
            "instructions": "Determine incident severity level",
            "criteria": {
                "SEV_1": "Critical revenue outage",
                "SEV_2": "Major feature degradation",
                "SEV_3": "Minor issue"
            }
        },
        "urgency": {
            "type": "score",
            "instructions": "Urgency rating",
            "criteria": ["Low", "Medium", "Immediate page"]
        },
        "rollback": {
            "type": "noul",
            "instructions": "Should a rollback be initiated?"
        }
    }
})

print(response["answers"])

Multimodal (Image) Evaluation

from PIL import Image

response = systemone(model, processor, {
    "model": "clef-flash",
    "state": "Review the uploaded invoice receipt.",
    "images": [Image.open("receipt.png")],
    "questions": {
        "legible": {"type": "noul", "instructions": "Is the receipt text clear and legible?"},
        "amount_exceeds_1000": {"type": "noul", "instructions": "Is total > $1000 USD?"}
    }
})

print(response["answers"])

Installation

pip install torch transformers huggingface_hub pillow

Usage (systemone API)

import sys
from huggingface_hub import snapshot_download

# Download repo and load custom joint schema model
repo_id = "Vishva007/clef-flash-W4A16-AutoRound"  # or auto-gptq / llm-compressor variant
model_path = snapshot_download(repo_id)
sys.path.insert(0, model_path)

from joint_schema_model import load_release_model, systemone

model, processor = load_release_model(model_path, device="cuda")

# 1. Text / JSON Decision Example
response = systemone(model, processor, {
    "model": "clef-flash",
    "state": "Prod database latency spiked to 4,000ms. Checkout failing with 504 Gateway Timeouts.",
    "questions": {
        "severity": {
            "type": "choice",
            "instructions": "Determine incident severity level",
            "criteria": {
                "SEV_1": "Critical revenue outage",
                "SEV_2": "Major feature degradation",
                "SEV_3": "Minor issue"
            }
        },
        "urgency": {
            "type": "score",
            "instructions": "Urgency rating",
            "criteria": ["Low", "Medium", "Immediate page"]
        },
        "rollback": {
            "type": "noul",
            "instructions": "Should a rollback be initiated?"
        }
    }
})

print(response["answers"])

Multimodal (Image) Evaluation

from PIL import Image

response = systemone(model, processor, {
    "model": "clef-flash",
    "state": "Review the uploaded invoice receipt.",
    "images": [Image.open("receipt.png")],
    "questions": {
        "legible": {"type": "noul", "instructions": "Is the receipt text clear and legible?"},
        "amount_exceeds_1000": {"type": "noul", "instructions": "Is total > $1000 USD?"}
    }
})

print(response["answers"])

---

Formats Available

---

🚀 Deploy on RunPod

One-click launch environments pre-configured with PyTorch, CUDA, and dependencies for fine-tuning or quantization.

🎁 Need GPU compute? Sign up via RunPod and get $5–$500 in free credits when you add your first $10.

PyTorch 2.15

PyTorch 2.14

PyTorch 2.13

PyTorch 2.12

---

Acknowledgements

• Base model by Cloudflare.
• Quantization powered by Intel AutoRound.