[tokinasin]

Overview

this PR is wip, requires #30206 and #30288

Apply YaRN truncation config to PLaMo-3 models.

(Theoretically, gpt-oss should also see performance improvements through similar fixes, but I haven't tested yet.)

Additional information

• KL divergence at the input tail before and after the fix

Approx. input length (tokens)	KL (before)	KL (after)
512	5.57e-4	2.81e-6
4,096	5.93e-2	5.84e-6
8,192	1.35e-1	6.69e-6
16,384	6.78e-1	9.60e-6

Experimental conditions

• Model: pfnet/plamo-3-610m-fin-instruct
• Input: System and user messages (Japanese financial-analysis question) with chat template applied, followed by reasoning tag (<|plamo:begin_think:plamo|>). Tested with 3 prompts per length.
• Measurement: The next-token probability distribution immediately after the reasoning tag.
• Reference: using FP32 on CPU inference with huggingface/transformers.
• llama.cpp: GGUF converted with --outtype f16 for both revisions, using CPU inference.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES
  • I used Codex for research, code reviews, etc.