Chranos/enginex-mlu370-vllm
1
0
Fork 0
forked from EngineX-Cambricon/enginex-mlu370-vllm
Code Pull Requests Activity

add deepseekv3 and llama4

Browse Source
This commit is contained in:
Chranos
2026-02-11 15:24:13 +08:00
parent 2ad23aa8da
commit c584139543
5 changed files with 320 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
vllm-v0.6.2/vllm/config.py
View File

@@ -1403,6 +1403,18 @@ class SpeculativeConfig:
draft_hf_config = draft_model_config.hf_config
# Detect DeepSeek V3 MTP: same model path with
# num_nextn_predict_layers > 0
num_nextn = getattr(draft_hf_config,
"num_nextn_predict_layers", 0)
if (num_nextn and num_nextn > 0
and getattr(draft_hf_config, "model_type", "")
in ("deepseek_v3",)):
draft_hf_config.model_type = "deepseek_mtp"
draft_hf_config.architectures = ["DeepSeekMTPModel"]
if num_speculative_tokens is None:
num_speculative_tokens = num_nextn
if (num_speculative_tokens is not None
and hasattr(draft_hf_config, "num_lookahead_tokens")):
draft_hf_config.num_lookahead_tokens = num_speculative_tokens
@@ -1421,7 +1433,7 @@ class SpeculativeConfig:
f"{num_speculative_tokens=} was provided.")
if enable_chunked_prefill and draft_hf_config.model_type in (
"medusa", "mlp_speculator", "eagle"):
"medusa", "mlp_speculator", "eagle", "deepseek_mtp"):
raise ValueError(
"Chunked prefill and hidden-state based draft models are "
"not compatible.")