vllm-cmp170hx-pp-int8 — custom vLLM build (PP2 + INT8 KV)

Source tarball of our patched vLLM fork for 2x CMP 170HX (GA100, sm80): pipeline-parallel PP2 (28,20 tiles), INT8 per-token-head KV cache, QSA sparse prefill, MTP speculative decoding, custom allreduce over PCIe P2P.

File: vllm-cmp170hx-pp-34a2085.tar.gz — git archive of HEAD 34a2085 (branch feature/int8-kv, base https://github.com/Morrowmake/vllm-cmp170hx) + runner scripts run_pp2_int8.sh, run_pp2.sh
sha256: dfcaf69cb75178f0da09c8d125279fc7a8412491e74d2ac8d16b208aa1e83122
Size: 45 574 609 bytes, 9064 files
Runtime: Python 3.13 venv, CUDA 13.2, editable install; launch recipe: see RECIPE_PP2_INT8_KV.md in SuperPUPS/Qwen3.8-Flash-Next-INT8x25-AttnBF16
Prod: qfn-vllm2-pp2-int8.service, :8000, max-num-seqs 16, ctx 262144

Uploaded 06.11.2026 by Dagaz (card t_9b83ef51).

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for SuperPUPS/vllm-cmp170hx-pp-int8

Base model

Qwen/Qwen3.8-Flash-Next

Finetuned

(117)

this model