[mottopanikeiku]

Nemotron3Renderer.render() raises ValueError: System message must be at the beginning. for any system message that is not the first message. The Nemotron 3 chat templates (Nano, Super, Ultra, 3.5 Lightning) accept that message and render it as its own turn through the user branch:

{%- elif message.role == "user" or message.role == "system" %}
    {{- '<|im_start|>' + message.role + '\n' }}

bridge_to_next_turn in the same renderer already renders an appended system message this way, so such a conversation can be extended through the bridge but not rendered in full. In verifiers' train client a tail that contains a system message is not bridged, so the full render() is the path that runs. verifiers' Responses dialect also turns instructions plus a developer input item into two system messages.

This change renders a non-leading system message through the user branch under its own role header, as the template does. The reasoning-effort hint still goes only on the last user message. The later-system parity scenario now also runs for nemotron-3, nemotron-3-ultra and nemotron-3.5.

Reproducer:

from renderers import create_renderer
from renderers.base import load_tokenizer

tok = load_tokenizer("nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16")
renderer = create_renderer(tok)
messages = [
    {"role": "system", "content": "You are a coding agent."},
    {"role": "user", "content": "Run the tests."},
    {"role": "assistant", "content": "All 12 tests pass."},
    {"role": "system", "content": "You have 5 turns left."},
    {"role": "user", "content": "Now add a test for the empty-input case."},
]
expected = list(
    tok.apply_chat_template(
        messages, add_generation_prompt=True, tokenize=True, return_dict=False
    )
)
print("template tokens:", len(expected))
got = renderer.render_ids(messages, add_generation_prompt=True)
print("renderer tokens:", len(got), "identical:", got == expected)

	before	after
reproducer above	ValueError	68 tokens, identical to the template
verifiers ResponsesDialect.parse_request of an instructions + developer request, rendered with the nemotron-3 config from configs/advanced/nemotron-3-super/swe.toml	ValueError	309 tokens, identical to the template (text parts joined first, as vLLM does)
pytest tests/test_parity.py -k "Nemotron and later-system"	32 failed	32 passed

After the change, pytest tests -k "nemotron or Nemotron" gives 1423 passed, 17 skipped (PIL/torch-dependent and capability skips). ruff 0.15.12 check and format pass on the changed files.

I found this by reading and testing the code.

I made this PR and directed its development, with AI assistance from Claude and GPT-6 Astra.

Note

Render later system messages in Nemotron3 renderer instead of raising

• Only the initial system message goes through the separate system-prompt path; later system messages now use the user-like emission path in nemotron3.py
• The emitted role header is derived from the actual message role, so later system messages are labeled as system rather than user
• Adds Nemotron-3, Nemotron-3 Ultra, and Nemotron-3.5 to the later-system scenario's renderer allowlist in parity.py
• Behavioral Change: conversations with non-initial system messages now render instead of being rejected

^{Macroscope summarized 83e2027.}