Tag: ascend
All the articles with the tag "ascend".
-
Deploying Qwen3.6-27B-W8A8 on Huawei Ascend 910B with vLLM Ascend
Updated:A field-tested deployment guide for Qwen3.6-27B-W8A8 on Huawei Ascend 910B/A2 with vLLM Ascend v0.19.1rc1 and vLLM v0.19.1, covering CANN 8.5.2, Python 3.11, torch-npu 2.9, Triton Ascend, native kernels, ATB, graph mode, and the failure fixes needed to make deployment repeatable.
-
How We Successfully Started Qwen3-Coder-Next on Huawei Ascend 910B with vLLM-Ascend 0.17
Complete deployment guide for Qwen3-Coder-Next on Huawei Ascend 910B with vLLM-Ascend 0.17. Documents the working configuration after multiple failed attempts, covering CANN 8.5.1, W8A8 quantization, worker startup modes, and KV cache management.
-
Mixed AI Compute Pools: Architecture Patterns for Heterogeneous GPU Infrastructure
A comprehensive guide to building production-ready mixed GPU pools. Explores capability-based segmentation, elastic burst patterns, multi-cloud federation, and real-world case studies including Huawei Ascend + NVIDIA hybrid deployments.
-
Deploying GLM-5 W4A8 on Huawei Cloud Ascend 910B x8
A field report on deploying GLM-5 W4A8 on ModelArts with 8 Ascend 910B cards. We solved Python stack issues and got the API server ready, but hit a decode blocker at the Ascend runtime layer. This is a practical summary of what worked, what broke, and where the real problem turned out to be.