From 0fff87a18044cdbf7af72157451ebdd801e3ff69 Mon Sep 17 00:00:00 2001 From: Paul Buetow Date: Wed, 18 Mar 2026 18:56:28 +0200 Subject: gpt-oss-120b: raise max_model_len to 163840 (160K) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 131K was still too small — observed 135K token conversations in practice. Physical KV capacity is 168K blocks so 160K is safe without OOM. Co-Authored-By: Claude Sonnet 4.6 (1M context) --- snippets/hyperstack/hyperstack-vm.toml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'snippets/hyperstack') diff --git a/snippets/hyperstack/hyperstack-vm.toml b/snippets/hyperstack/hyperstack-vm.toml index d7d09e3..cd16615 100644 --- a/snippets/hyperstack/hyperstack-vm.toml +++ b/snippets/hyperstack/hyperstack-vm.toml @@ -124,12 +124,12 @@ tool_call_parser = "" # OpenAI GPT-OSS 120B — powerful MoE (5.1B active / 117B total, MXFP4), ~65 GB on A100. # MXFP4 KV cache is very compact: vLLM allocates 168K token blocks (10560×16) at 0.92 util. -# 131072 fits Claude Code's ~33K system prompt with room for long conversations. +# 163840 (160K) stays within physical KV capacity; handles long Claude Code sessions (135K+). # tool_call_parser = "" disables --enable-auto-tool-choice (same reason as gpt-oss-20b). [vllm.presets.gpt-oss-120b] model = "openai/gpt-oss-120b" container_name = "vllm_gpt_oss_120b" -max_model_len = 131072 +max_model_len = 163840 gpu_memory_utilization = 0.92 tensor_parallel_size = 1 tool_call_parser = "" -- cgit v1.2.3