Qwen3.6 27B KV cache experiment

Measured Qwen3.6-27B-Q5_K_M with llama-perplexity.exe on a 3090 eGPU with a 200k context setup.

• Without KV cache quantization, it recorded PPL 6.9233 ± 0.04564.
• Repeated measurements were done with settings such as --cache-type-k turbo3 --cache-type-v turbo3, and the author noted that several quants generally fell within an acceptable range.
• Since the tools and settings used are provided together, this can be viewed as a reproducible benchmark example for the same model and environment.
• However, based on the author's personal experience, the turbo3 cache didn't work well on a 35B-class model, and they thought smaller models might be more sensitive to excessive V cache compression.