Upload tools/nemotron_fmt.py with huggingface_hub

tools/nemotron_fmt.py

@@ -0,0 +1,84 @@
+"""Chat rendering for NVIDIA-Nemotron-3.5-Lightning-30B-A3B.
+
+Same contract as ``glimmer_fmt``: ``render()`` turns a pool dialogue record into
+the exact byte string the model sees at inference.
+
+Unlike ``glimmer_fmt`` this is not a hand port. Nemotron's ``chat_template.jinja``
+is driven directly through ``transformers.apply_chat_template``, which is the
+reference implementation, so byte equality is true by construction rather than
+by assertion.
+
+Differences from Muse-Glimmer that matter for the corpus:
+  * special tokens are ``<|im_start|>`` / ``<|im_end|>`` / ``<think>`` /
+    ``<tool_call>``, not ``<|start|>`` / ``<|message|>`` / ``<|eot|>``
+  * ``add_bos_token`` is false in tokenizer_config, so documents carry no BOS
+  * the template takes no date arguments, so builds are reproducible with no
+    pinning needed (Glimmer needed current_date to avoid strftime_now)
+  * there is no reasoning_strength dial; the template has a boolean
+    ``enable_thinking``, so the pool's four strengths collapse onto it
+"""
+from __future__ import annotations
+
+import os
+from typing import Any, Sequence
+
+# Kept for interface parity with glimmer_fmt; Nemotron's template uses neither.
+DEFAULT_REASONING = "high"
+DEFAULT_KNOWLEDGE_CUTOFF = None
+
+MODEL_DIR = os.environ.get("NEMOTRON_MODEL_DIR", "/root/work/nemo-bf16")
+
+# Reasoning strengths in the pool that mean "do not emit a thinking channel".
+_NO_THINK = {"none", "off", "no", "disabled", "low0"}
+
+_tok = None
+
+
+def _tokenizer():
+    global _tok
+    if _tok is None:
+        from transformers import AutoTokenizer  # heavy; imported lazily
+        _tok = AutoTokenizer.from_pretrained(MODEL_DIR, trust_remote_code=False)
+        if not _tok.chat_template:
+            raise SystemExit(f"no chat_template found in {MODEL_DIR}")
+    return _tok
+
+
+def _clean(messages: Sequence[dict]) -> list[dict]:
+    """Drop Glimmer-only per-message keys the Nemotron template does not read.
+
+    ``recipient``/``end_turn``/``channel`` are Muse-Glimmer routing fields.
+    ``reasoning_content`` and ``tool_calls`` are kept - Nemotron's template
+    consumes both.
+    """
+    keep = {"role", "content", "reasoning_content", "tool_calls",
+            "tool_call_id", "name"}
+    out = []
+    for m in messages:
+        d = {k: v for k, v in m.items() if k in keep and v is not None}
+        d.setdefault("content", "")
+        out.append(d)
+    return out
+
+
+def render(messages: Sequence[dict],
+           tools: Sequence[dict] | None = None,
+           reasoning_strength: str | None = None,
+           knowledge_cutoff: Any = None,      # unused, interface parity
+           current_date: Any = None,          # unused, interface parity
+           namespace_descriptions: Any = None,  # unused, Glimmer-only
+           add_generation_prompt: bool = False,
+           add_bos: bool = False) -> str:
+    tok = _tokenizer()
+    enable_thinking = str(reasoning_strength or DEFAULT_REASONING).lower() not in _NO_THINK
+
+    text = tok.apply_chat_template(
+        _clean(messages),
+        tools=list(tools) if tools else None,
+        enable_thinking=enable_thinking,
+        add_generation_prompt=add_generation_prompt,
+        tokenize=False,
+    )
+    if add_bos and tok.bos_token and not text.startswith(tok.bos_token):
+        text = tok.bos_token + text
+    return text