|
|
@@ -70,12 +70,6 @@ DCP_SIZE = "1"
|
|
|
# empty revision is ignored.
|
|
|
MODEL_PATH = "/opt/local-hf/hub/models--deepseek-ai--DeepSeek-V4-Flash-Vision-Exp/snapshots/e46e16bf6035c6f317eb2ac7458eb0362926d402"
|
|
|
MODEL_ID = "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp"
|
|
|
-DSPARK_# his script appends --revision "<empty>" unconditionally, which breaks
|
|
|
-# offline snapshot resolution (revision "" matches nothing). Point MODEL_PATH
|
|
|
-# at the cached snapshot path directly — vllm uses the existing path and the
|
|
|
-# empty revision is ignored.
|
|
|
-MODEL_PATH = "/opt/local-hf/hub/models--deepseek-ai--DeepSeek-V4-Flash-Vision-Exp/snapshots/e46e16bf6035c6f317eb2ac7458eb0362926d402"
|
|
|
-MODEL_ID = "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp"
|
|
|
SERVED_MODEL_NAME = "deepseek-v4-vision"
|
|
|
GPU_MEMORY_UTILIZATION = "0.80"
|
|
|
KV_CACHE_MEMORY_BYTES = "12884901888"
|