qwen3-4b-awq@0.1.0

Immutable release / 0.1.0

PlannedPlanning blockedNot measured

Qwen3-4B · AWQ W4A16

A pinned AWQ W4A16 recipe targeting compressed-tensors export and vLLM inference. Dataset selection remains an explicit planning blocker.

README.md

Before you use this recipe

The execution recipe is designed, but immutable calibration and evaluation datasets have not been selected. Planning fails closed until both dataset references are pinned.

No quantized artifact has been produced or published. No accuracy, latency, throughput, memory, or hardware evidence bundle exists.

Model

Model	`Qwen/Qwen3-4B`
Family	Qwen3
Revision	`1cfa9a7208912126459214e8b04321603b3df60c`

Compression method

This recipe specifies AWQ with 4-bit weights and 16-bit activations (W4A16), using a group size of 128. The export format is compressed-tensors, targeting vLLM inference.

Ordered stages

quantize
Recipe plugin
lazybrick.plugin/awq@0.1.0
Upstream implementation
https://github.com/vllm-project/llm-compressorde46bfd53513aa87571a8b056a06aeaa5da1c69c
Stage parameters
```
{
  "weight_bits": 4,
  "group_size": 128,
  "ignore": [
    "lm_head"
  ]
}
```

These pins identify authored declarations. The Hub has not verified the adapter package, installed source, or execution binding.

Numerical semantics

quantize: Unspecified

This recipe pins an implementation but does not declare a numerical profile.

Conformance: Not verified

A declaration is not evidence that an implementation, artifact, or runtime conforms.

Requirements

Accelerator	NVIDIA CUDA
Declared memory target	40 GiB
Runtime	vLLM
Validator	lazybrick 0.0.1
Identity command	`lazybrick fingerprint`

Compatibility is declared, not measured.

Calibration and evaluation

Calibration data	`hf-dataset://REPLACE-ME/calibration-set` `REPLACE-ME`
Calibration samples	256
Evaluation data	`hf-dataset://REPLACE-ME/held-out-set` `REPLACE-ME`
Evaluation protocol	`token-loss-v1`
Evaluation sample limit	128

Protocol and sample limits describe the intended workload; no measured result is implied.

Use this recipe

Download qwen3-4b-awq-0.1.0.recipe.yaml using the download button above. The Usage tab has copyable CLI and Python examples for this exact file. Review the Blockers tab before running it, and inspect or copy the full source in Recipe YAML.

Recipe identity

Recipe	qwen3-4b-awq
Version	0.1.0
Recipe digest	`d375e4452dc0b86bae059c6f7b5d38d8ae3d40f1c31216156169c07aa259194f`
File SHA-256	`3b6ccf40e164c895e31463a1ab1933ba6a886e13356a52ff8770ea7f965f00bc`

The recipe digest identifies canonical authored content; the file hash identifies the exact downloaded bytes. Neither is a plan, artifact, run, or evidence identity. Presentation metadata is available separately in entry.json.

Validation

Blocking decisions

calibration.datasetmutable_reference

Calibration dataset URI and revision are placeholders.

Choose a license-compatible dataset and pin its immutable revision at calibration.dataset in a new recipe version. Preserve this release unchanged.

evaluation.datasetmutable_reference

Evaluation dataset URI and revision are placeholders.

Choose a license-compatible dataset and pin its immutable revision at evaluation.dataset in a new recipe version. Preserve this release unchanged.

Usage

Use the exact recipe

These examples use the downloaded file qwen3-4b-awq-0.1.0.recipe.yaml.

Install the recorded validator, then run these commands in the download directory. Installation may access the package registry; validation and fingerprinting run locally without model downloads.

Install CLI · uv

uv tool install 'lazybrick==0.0.1'

For the Python example, use uv add 'lazybrick==0.0.1' in your project. With pip, use python -m pip install 'lazybrick==0.0.1'.

CLI · schema validation

lazybrick validate qwen3-4b-awq-0.1.0.recipe.yaml

Expected: exit code 0 and a valid-recipe message containing the digest below. This checks the published validator’s schema boundary; it does not resolve references, validate execution compatibility, or clear planning blockers.

Authored recipe identity

lazybrick fingerprint qwen3-4b-awq-0.1.0.recipe.yaml

Expected SHA-256: d375e4452dc0b86bae059c6f7b5d38d8ae3d40f1c31216156169c07aa259194f

The recorded lazybrick 0.0.1 release provides validation and fingerprinting. Planning requires a separately reviewed core checkout, target and plugin manifests, immutable inputs, and verified execution bindings. Review the planning blockers before preparing execution.

Python

from lazybrick import load_recipe

recipe = load_recipe("qwen3-4b-awq-0.1.0.recipe.yaml")
print(recipe.digest)

Source

Authored recipe

qwen3-4b-awq-0.1.0.recipe.yaml

# LazyBrick recipe, schema v0.1.
#
# The model and the plugin implementation are pinned to real immutable commits.
# The calibration dataset is NOT: choosing a licence-compatible calibration set
# is still an open blocking decision, so its URI is a placeholder and `lazybrick
# plan` will refuse this recipe until it is replaced.

schema_version: "0.1"

model:
  uri: hf://Qwen/Qwen3-4B
  revision: 1cfa9a7208912126459214e8b04321603b3df60c
  trust_remote_code: false

stages:
  - id: quantize
    plugin: lazybrick.plugin/awq
    plugin_version: "0.1.0"
    implementation:
      git: https://github.com/vllm-project/llm-compressor
      commit: de46bfd53513aa87571a8b056a06aeaa5da1c69c
    parameters:
      weight_bits: 4
      group_size: 128
      ignore:
        - lm_head

calibration:
  dataset:
    uri: hf-dataset://REPLACE-ME/calibration-set
    revision: REPLACE-ME
  samples: 256
  seed: 42
  preprocessing_id: qwen3-chat-template-v1
  max_sequence_length: 2048

export:
  format: compressed-tensors
  runtime: vllm

target:
  accelerator_family: nvidia-cuda
  # A string, not a number: 8.0 and 8.00 must not collapse to the same value.
  min_compute_capability: "8.0"
  device_count: 1
  min_memory_gib: 40
  runtime: vllm

evaluation:
  protocol_id: token-loss-v1
  dataset:
    uri: hf-dataset://REPLACE-ME/held-out-set
    revision: REPLACE-ME
  max_samples: 128
  seed: 1234