Update entrypoint.sh
Browse files- entrypoint.sh +3 -1
entrypoint.sh
CHANGED
|
@@ -31,5 +31,7 @@ exec /app/llama-server \
|
|
| 31 |
--cache-type-k "$CACHE_TYPE_K" \
|
| 32 |
--cache-type-v "$CACHE_TYPE_V" \
|
| 33 |
--reasoning "$REASONING_MODE" \
|
| 34 |
-
-fa "$FLASH_ATTN"
|
|
|
|
|
|
|
| 35 |
|
|
|
|
| 31 |
--cache-type-k "$CACHE_TYPE_K" \
|
| 32 |
--cache-type-v "$CACHE_TYPE_V" \
|
| 33 |
--reasoning "$REASONING_MODE" \
|
| 34 |
+
-fa "$FLASH_ATTN" \
|
| 35 |
+
--spec-type draft-mtp \
|
| 36 |
+
--spec-draft-n-max 4
|
| 37 |
|