Upload folder using huggingface_hub
Browse files
build.sh
ADDED
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
|
3 |
+
INFERENCE_PRECISION=float16
|
4 |
+
MAX_BEAM_WIDTH=4
|
5 |
+
MAX_BATCH_SIZE=64
|
6 |
+
|
7 |
+
checkpoint_dir=$1
|
8 |
+
output_dir=$2
|
9 |
+
|
10 |
+
trtllm-build --checkpoint_dir ${checkpoint_dir}/encoder \
|
11 |
+
--output_dir ${output_dir}/encoder \
|
12 |
+
--moe_plugin disable \
|
13 |
+
--enable_xqa disable \
|
14 |
+
--max_batch_size ${MAX_BATCH_SIZE} \
|
15 |
+
--gemm_plugin disable \
|
16 |
+
--bert_attention_plugin ${INFERENCE_PRECISION} \
|
17 |
+
--max_input_len 3000 --max_seq_len=3000
|
18 |
+
|
19 |
+
|
20 |
+
trtllm-build --checkpoint_dir ${checkpoint_dir}/decoder \
|
21 |
+
--output_dir ${output_dir}/decoder \
|
22 |
+
--moe_plugin disable \
|
23 |
+
--enable_xqa disable \
|
24 |
+
--max_beam_width ${MAX_BEAM_WIDTH} \
|
25 |
+
--max_batch_size ${MAX_BATCH_SIZE} \
|
26 |
+
--max_seq_len 114 \
|
27 |
+
--max_input_len 14 \
|
28 |
+
--max_encoder_input_len 3000 \
|
29 |
+
--gemm_plugin ${INFERENCE_PRECISION} \
|
30 |
+
--bert_attention_plugin ${INFERENCE_PRECISION} \
|
31 |
+
--gpt_attention_plugin ${INFERENCE_PRECISION}
|