Download mtp_draft/plugin/q38_mtp_draft.py from ctech123/ctech-fle-gt: direct link, hf CLI and curl.
- Browser
- Download file 807 Bytes
-
https://huggingface.co/ctech123/ctech-fle-gt/resolve/main/mtp_draft/plugin/q38_mtp_draft.py
- Command line
-
hf download hf://ctech123/ctech-fle-gt/mtp_draft/plugin/q38_mtp_draft.py
-
curl -L -o q38_mtp_draft.py https://huggingface.co/ctech123/ctech-fle-gt/resolve/main/mtp_draft/plugin/q38_mtp_draft.py
807 Bytes
| """vLLM general plugin: install the cheap Qwen3.8-Flash-Next MTP drafter (q38_draft.py) in every vLLM process. | |
| Inactive unless Q38_DRAFT_VOCAB is set, so the same environment still runs unpatched servers. | |
| Q38_DRAFT_HOME directory holding q38_draft.py / nvfp4_ops.py (this repo's mtp_draft/) | |
| Q38_DRAFT_VOCAB draft vocabulary mask (.pt bool[248320]); empty -> plugin inactive | |
| Q38_DRAFT_HEAD_Q marlin (default) | cutlass | bf16 draft lm_head | |
| Q38_DRAFT_MTP_Q marlin (default) | cutlass | bf16 MTP block BF16 linears | |
| """ | |
| import os | |
| import sys | |
| def main(): | |
| if not os.environ.get("Q38_DRAFT_VOCAB"): | |
| return | |
| home = os.environ.get("Q38_DRAFT_HOME") | |
| if home and home not in sys.path: | |
| sys.path.insert(0, home) | |
| import q38_draft | |
| q38_draft.install_from_env() | |