Update entrypoint.sh
Browse files- entrypoint.sh +2 -4
entrypoint.sh
CHANGED
|
@@ -1,10 +1,10 @@
|
|
| 1 |
#!/bin/sh
|
| 2 |
set -eu
|
| 3 |
|
| 4 |
-
MODEL_SPEC="${MODEL_SPEC:-openbmb/MiniCPM-V-4.7-1B-gguf:
|
| 5 |
HOST="${HOST:-0.0.0.0}"
|
| 6 |
PORT="${PORT:-7860}"
|
| 7 |
-
CTX_SIZE="${CTX_SIZE:-
|
| 8 |
THREADS="${THREADS:-2}"
|
| 9 |
PARALLEL="${PARALLEL:-1}"
|
| 10 |
REASONING_MODE="${REASONING_MODE:-off}"
|
|
@@ -35,5 +35,3 @@ exec /app/llama-server \
|
|
| 35 |
--reasoning "$REASONING_MODE" \
|
| 36 |
-fa "$FLASH_ATTN" \
|
| 37 |
$SPEC_FLAGS
|
| 38 |
-
|
| 39 |
-
|
|
|
|
| 1 |
#!/bin/sh
|
| 2 |
set -eu
|
| 3 |
|
| 4 |
+
MODEL_SPEC="${MODEL_SPEC:-openbmb/MiniCPM-V-4.7-1B-gguf:Q8_0}"
|
| 5 |
HOST="${HOST:-0.0.0.0}"
|
| 6 |
PORT="${PORT:-7860}"
|
| 7 |
+
CTX_SIZE="${CTX_SIZE:-16384}"
|
| 8 |
THREADS="${THREADS:-2}"
|
| 9 |
PARALLEL="${PARALLEL:-1}"
|
| 10 |
REASONING_MODE="${REASONING_MODE:-off}"
|
|
|
|
| 35 |
--reasoning "$REASONING_MODE" \
|
| 36 |
-fa "$FLASH_ATTN" \
|
| 37 |
$SPEC_FLAGS
|
|
|
|
|
|