Modelpublic

zhouxzh/ascend310b-llm-om-zoo

sourceHugging Faceotherupdated 25d agoView on Hugging Face

0likes

Model Card

Ascend 310B LLM OM Zoo

This release contains a Qwen2.5-0.5B-Instruct static-KV ONNX graph and Ascend OM artifacts tested with native ACL on two Ascend 310B hardware tiers. The repository is an experimental conversion and validation archive, not an official Huawei or Qwen release.

Included release

Variant	Hardware target	Artifact	SHA-256
B4	Ascend310B4 / 8T	`models/qwen2.5-0.5b-instruct-static-kv-1024/ascend310b4-cann8.0/qwen25-static-kv-1024-v2.om`	`f6650e52ff3908288763ef7957832ade606b0e554fa8fde986932f1ca1140eb8`
B1	Ascend310B1 / 20T	`models/qwen2.5-0.5b-instruct-static-kv-1024/ascend310b1-cann8.0/qwen25-static-kv-1024-b1.om`	`6bca884fbce746efdb02f8c9294cad5b2faa6c8b96cac9ec8c83730126298609`
Common	ONNX static-KV graph	`models/qwen2.5-0.5b-instruct-static-kv-1024/common/qwen25-static-kv-1024-v2.onnx`	`b4870df5da9c8cbef4163ceb65d4dc13433f2fd8ed5d2083ef3223d07d1a3c0e`

Exact byte sizes, contracts, lock files, and provenance are in manifests/artifact-manifest.json and manifests/SHA256SUMS.txt.

Layout

text

models/qwen2.5-0.5b-instruct-static-kv-1024/
  common/                 # ONNX, tokenizer and graph contract
  ascend310b4-cann8.0/    # B4 / 8T OM and descriptor contract
  ascend310b1-cann8.0/    # B1 / 20T OM and descriptor contract
licenses/qwen/LICENSE     # upstream Apache-2.0 text
manifests/                # hashes and sanitized provenance

The OM files require the matching Ascend SoC/CANN runtime and the native ACL service. They are not interchangeable conversion claims for arbitrary Ascend devices. Use the B4 and B1 variants according to the descriptor and lock files; do not infer compatibility from file names alone.

Upstream and license

The base model is Qwen/Qwen2.5-0.5B-Instruct. Its upstream Apache-2.0 text is preserved under licenses/qwen/LICENSE; see UPSTREAM.md for the pinned source and conversion provenance. This repository does not relicense upstream model artifacts. CANN/ACL redistribution terms and local deployment policy must be checked separately before commercial redistribution.

Limitations

The artifacts are experimental. The B4 result is the current 8T validation line. The B1 result is a separate 20T/dirty-base evidence line. Reported latency and Chinese quality are hardware- and revision-specific; they are not claims of official support or production stability.

TinyLlama historical artifact (blocked)

This repository also contains a separately archived TinyLlama 1.1B community OM and matching tokenizer under models/tinyllama-1.1b-chat/. It is published for reproducibility only. The OM/API experiment ran on Ascend310B4, but the Case9 long-output gate failed at max_tokens=32/48 with U+FFFD, the Chinese machine-quality result was 7/10, and current CANN 8.0 ATC evidence is not available. The source checkpoint is uploaded under source/ for provenance and must not be confused with the prebuilt OM. Keep this artifact blocked; do not use it as a production Chinese assistant or XiaoZhi backend. See validation/tinyllama-validation-summary.json and manifests/tinyllama-artifact-manifest.json for hashes and evidence.