Active active in llama-swap.yaml
| Run | Score | tok/step | R5 | Log |
|---|---|---|---|---|
| No results-*.log found for this id/alias. | ||||
"omni":
# THE ONLY MODEL ON THIS BOX THAT CAN HEAR. Qwen3-Omni-30B-A3B AWQ, 25.7 GiB,
# the largest model served here (15.2 of 16.3 GiB per card). Verified
# 2026-08-10: espeak-ng speech in -> 8/8 words back verbatim. vLLM reports
# Supported tasks: ['generate','transcription']. Every other model in the HF
# cache has ZERO audio tensors. See auto-memory qwen3-omni-audio-works.
#
# AUDIO NEEDS librosa/soundfile/av IN THE VENV, and the server must be
# started AFTER they are installed — vLLM caches a placeholder for a missing
# module and then 500s with "PlaceholderModule should not be used...".
# Install with /home/giovanni/vllm-env/bin/pip, NOT the shell's pip (that one
# is ~/.local/bin/pip and targets SYSTEM python). Never `pip install
# vllm[audio]`: it can upgrade vLLM and silently revert the MLA fp8 patch.
#
# 65536 is its NATIVE max_position_embeddings and fits easily — measured
# 3.86 GiB KV per GPU (168,464 tokens) against the 1.50 GiB total that 65k
# needs. 132 tok/s. Grill 42/46 with the LOWEST spread in the table (3), and
# 4/16 KLayout from memory, 2nd only to fable-fusion.
#
# ALWAYS give it the openpyxl reference for spreadsheet work: office is 0/18
# unassisted with FOUR 16k runaways, and 18/18 with the reference and none.
cmd: |
/usr/bin/env
HF_HOME=/home/giovanni/.cache/huggingface
CUDA_DEVICE_ORDER=PCI_BUS_ID
PATH=/home/giovanni/vllm-env/bin:/home/giovanni/.local/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
${vllm_bin} serve cyankiwi/Qwen3-Omni-30B-A3B-Instruct-AWQ-4bit
--served-model-name omni
--tensor-parallel-size 2
--host 127.0.0.1 --port 9136
--max-model-len 65536
--max-num-seqs 8
--gpu-memory-utilization 0.93
--kv-cache-dtype fp8
--override-generation-config '{"temperature":0.6}'
--enable-auto-tool-choice
--tool-call-parser hermes
proxy: http://127.0.0.1:9136