EngineX/xc-llm-kunlun
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[Feature] support deepseek v3/r1/v3.2 (#78)

Browse Source 
* [Feature] support deepseek v3/r1/v3.2

* fix gpt_oss

* update readme

* update readme

---------

Co-authored-by: hanhaowen <hanhaowen@baidu.com>
This commit is contained in:
baoqian426
2026-01-05 22:55:35 +08:00
committed by GitHub
parent 07bc24a555
commit ee0f50e68f
27 changed files with 5760 additions and 621 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
vllm_kunlun/platforms/kunlun.py
View File

@@ -177,6 +177,8 @@ class KunlunPlatform(Platform):
# if `VLLM_ATTENTION_BACKEND` is not set and we are using MLA, then
# we default to FlashMLA backend, so we need to force the blocksize
# here
use_sparse = hasattr(vllm_config.model_config.hf_config,
"index_topk")
use_flashmla = (envs.VLLM_ATTENTION_BACKEND is None \
or envs.VLLM_ATTENTION_BACKEND == "FLASHMLA")
from vllm.attention.ops.flashmla import is_flashmla_supported
@@ -185,6 +187,11 @@ class KunlunPlatform(Platform):
cache_config.block_size = 64
logger.info(
"Forcing kv cache block size to 64 for FlashMLA backend.")
if use_sparse and cache_config.block_size != 64:
cache_config.block_size = 64
logger.info(
"Forcing kv cache block size to 64 for FlashMLASparse "
"backend.")
if (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput"
and parallel_config.data_parallel_size > 1
@@ -224,6 +231,14 @@ class KunlunPlatform(Platform):
Returns:
str: Class name of the attention backend.
"""
if use_mla:
if use_sparse:
logger.info_once("Using Sparse MLA backend on V1 engine.")
# return ("vllm.v1.attention.backends.mla.flashmla_sparse."
# "FlashMLASparseBackend")
return ("vllm_kunlun.v1.attention.backends.mla.flashmla_sparse."
"FlashMLASparseBackend")
return "vllm_kunlun.v1.attention.backends.mla.flashmla.FlashMLABackend"
if use_v1:
return "vllm_kunlun.v1.attention.backends.kunlun_attn.KunlunAttentionBackend"
elif not use_mla: