diff options
Diffstat (limited to 'snippets/hyperstack')
| -rw-r--r-- | snippets/hyperstack/hyperstack-vm.toml | 8 |
1 files changed, 4 insertions, 4 deletions
diff --git a/snippets/hyperstack/hyperstack-vm.toml b/snippets/hyperstack/hyperstack-vm.toml index e739d5f..4e47a76 100644 --- a/snippets/hyperstack/hyperstack-vm.toml +++ b/snippets/hyperstack/hyperstack-vm.toml @@ -94,16 +94,16 @@ tool_call_parser = "qwen3_coder" # Requires trust_remote_code=true for the nemotron_h architecture. # Note: cyankiwi AWQ has model_type="nemotron_nas" (underscore); vLLM keys on "nemotron-nas" # (hyphen), so vLLM may not recognise it without trust_remote_code and latest vLLM. -# Tool calling: Nemotron uses a custom XML format (<tool_call><function=...><parameter=...>) -# not supported by any vLLM 0.17.1 built-in parser. tool_call_parser="" disables tool calling. -# Use for long-context analysis and reasoning; switch to qwen3-coder-next for agentic work. +# NVIDIA Nemotron-3-Super uses the same XML tool call format as Qwen3 XML: +# <tool_call><function=name><parameter=p>value</parameter></function></tool_call> +# qwen3_xml handles this format and is compatible with Nemotron's chat template. [vllm.presets.nemotron-super] model = "cyankiwi/NVIDIA-Nemotron-3-Super-120B-A12B-AWQ-4bit" container_name = "vllm_nemotron_super" max_model_len = 65536 gpu_memory_utilization = 0.92 tensor_parallel_size = 1 -tool_call_parser = "" +tool_call_parser = "qwen3_xml" trust_remote_code = true # nemotron_v3 reasoning parser exposes <think> tokens as reasoning_content in the API. extra_vllm_args = ["--reasoning-parser", "nemotron_v3"] |