vllm-ascend-quant-ext

Runtime-only extension for loading and executing versioned Ascend quantized model artifacts. It contains no calibration datasets, ModelSlim invocation, model conversion, PPL evaluation, or benchmark orchestration.

Installation alone changes no vLLM behavior. Admission is explicit and fail-closed:

vllm-ascend-quant-ext check --model /path/to/model
vllm-ascend-quant-ext plan --model /path/to/model
vllm-ascend-quant-ext render --model /path/to/model

plan and render describe an explicit direct-diagnostic path. They do not produce Extension Manager activation evidence. The plugin validates the model contract before importing vLLM-Ascend implementation modules and never modifies the model directory.

This extension is for model-weight quantization. It does not own KV-cache format, allocation, compression, scheduling, or the lifecycle of vllm-ascend-adaptive-quantized-kv-hust.

The proposed Manager adapter is exposed as a pure-data symbol:

from vllm_ascend_quant_ext.manager import provider

provider.descriptor()
provider.check({"enabled": True, "model": "/path/to/model"})
provider.plan({"enabled": True, "model": "/path/to/model"})
provider.render({"enabled": True, "model": "/path/to/model"})

This adapter is import-only: inspection and artifact checking are supported, but enabled plan and render fail closed until vLLM-Ascend publishes the declared typed Host protocols. See ../docs/extension-manager-handoff.md.

Extension Bundle identity

PyPI distribution: vllm-ascend-quant-ext
Python package:     vllm_ascend_quant_ext
Bundle ID:          org.vllm-hust.ascend-quant-runtime
Component ID:       ascend-quant-artifact-validator
Full component ID:  org.vllm-hust.ascend-quant-runtime/ascend-quant-artifact-validator
Contract:           vllm.ascend.quantized-artifact-loader.v1
Execution planes:   worker, device
Status:             import_only

The wheel exposes two distinct entry points:

• vllm_hust.extension_bundles discovers static Bundle metadata;
• vllm.general_plugins provides an explicit direct-diagnostic registration path.

The first is the Manager discovery target. The Bundle does not automatically select the second entry point or inject environment/additional-config values. That prevents partial integration from starting an unsafe legacy path. Formal activation remains blocked until vLLM-Ascend accepts and materializes both declared Host protocols.

The namespaced ASCEND_QUANT_W8A8 registration delegates to the selected vLLM-Ascend W8A8 linear and MoE implementations. The extension owns admission and alias registration, while vLLM-Ascend continues to own weight loading, parameter layout, execution-role selection, and NPU operators.

The Bundle manifest is located at:

vllm_ascend_quant_ext/manifests/vllm-hust-extension-v0.2.json

The identifiers use the current 0.2-experimental Manager baseline. This is still an alpha/NO-GO integration: import_only must not be changed to an active status before the Extension Manager and vLLM-Ascend owners approve the Host loader/operator contracts.