compare with original vllm mtp spec?
is there results of comparing with original vllm mtp spec with num_speculative_tokens=5
Under the same GSM8K evaluation settings, GLM-5.2’s native MTP model achieves an average accepted length of approximately 2.9 and a token acceptance rate of approximately 0.37. In comparison, RedHatAI/GLM-5.2-speculator.dspark achieves an average accepted length of only around 1.4 and a token acceptance rate of around 0.17.
Based on these results, the speculative decoding acceptance performance of RedHatAI/GLM-5.2-speculator.dspark appears to be significantly lower than that of GLM-5.2’s native MTP model. Is this expected behavior, or are there any additional configurations, specific chat templates, or matching inference parameters required to achieve the expected performance?
i am also having quite bad acceptance rate, not sure if it is based on the FP8 version