biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Deepseek v3 (#11502)
Some checks failed
Create Release / Create Release (push) Has been cancelled
Details

Browse Source 
Signed-off-by: mgoin <michael@neuralmagic.com>
Co-authored-by: mgoin <michael@neuralmagic.com>
Co-authored-by: robertgshaw2-neuralmagic <rshaw@neuralmagic.com>
This commit is contained in:
Simon Mo
2024-12-26 16:09:44 -08:00
committed by GitHub
parent 55fb97f7bd
commit f49777ba62
7 changed files with 886 additions and 60 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
vllm/config.py
View File

@@ -596,6 +596,12 @@ class ModelConfig:
self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
self.max_model_len)
if (self.hf_config.model_type == 'deepseek_v3'
and not self.enforce_eager):
logger.warning("CUDA graph is not supported for Deepseek V3 yet, "
"fallback to the eager mode.")
self.enforce_eager = True
def _verify_bnb_config(self) -> None:
"""
The current version of bitsandbytes (0.44.0) with 8-bit models does not
@@ -712,8 +718,9 @@ class ModelConfig:
def get_head_size(self) -> int:
# TODO remove hard code
if hasattr(self.hf_text_config, "model_type"
) and self.hf_text_config.model_type == 'deepseek_v2':
if hasattr(self.hf_text_config,
"model_type") and (self.hf_text_config.model_type
in ('deepseek_v2', 'deepseek_v3')):
# FlashAttention supports only head_size 32, 64, 128, 256,
# we need to pad head_size 192 to 256
return 256