GitLab

[Feature] [Spec decode]: Enable MLPSpeculator/Medusa and `prompt_logprobs` with ChunkedPrefill (#10132)
Signed-off-by: NickLucche <nlucches@redhat.com>
Signed-off-by: wallashss <wallashss@ibm.com>
Co-authored-by: wallashss <wallashss@ibm.com>

add VLLM_OPTEST_MODELS_PATH/OPTEST_MODELS_PATH  to load models from local path instead of Hugging Face Hub

Removing the block manager v1. This is the initial piece of prefix-caching-centric design. In order to achieve prefix-caching-centric design, we need to simplify the code path so that we only use v2 block manager (which has much higher performance on prefix caching).

Signed-off-by: Travis Johnson <tsjohnso@us.ibm.com>

Co-authored-by: youkaichao <youkaichao@126.com>

Signed-off-by: Travis Johnson <tsjohnso@us.ibm.com>

[Spec Decode] Disable Log Prob serialization to CPU for spec decoding for both draft and target models. (#6485)