Upload folder using huggingface_hub
Browse files- build_whisper_fp16.sh +34 -0
- build_whisper_int8.sh +37 -0
- tllm_checkpoint/decoder/config.json +38 -0
- tllm_checkpoint/decoder/rank0.safetensors +3 -0
- tllm_checkpoint/encoder/config.json +15 -0
- tllm_checkpoint/encoder/rank0.safetensors +3 -0
- tllm_checkpoint_int8/decoder/config.json +38 -0
- tllm_checkpoint_int8/decoder/rank0.safetensors +3 -0
- tllm_checkpoint_int8/encoder/config.json +15 -0
- tllm_checkpoint_int8/encoder/rank0.safetensors +3 -0
build_whisper_fp16.sh
ADDED
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
export CUDA_VISIBLE_DEVICES=1
|
2 |
+
INFERENCE_PRECISION=float16
|
3 |
+
MAX_BEAM_WIDTH=4
|
4 |
+
MAX_BATCH_SIZE=8
|
5 |
+
checkpoint_dir=tllm_checkpoint
|
6 |
+
output_dir=distill_whisper_large_v3
|
7 |
+
|
8 |
+
|
9 |
+
# trtllm-build --checkpoint_dir ${checkpoint_dir}/encoder \
|
10 |
+
# --output_dir ${output_dir}/encoder \
|
11 |
+
# --paged_kv_cache disable \
|
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 |
+
# --remove_input_padding disable
|
18 |
+
|
19 |
+
# trtllm-build --checkpoint_dir ${checkpoint_dir}/decoder \
|
20 |
+
# --output_dir ${output_dir}/decoder \
|
21 |
+
# --paged_kv_cache disable \
|
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 100 \
|
27 |
+
# --max_input_len 14 \
|
28 |
+
# --max_encoder_input_len 1500 \
|
29 |
+
# --gemm_plugin ${INFERENCE_PRECISION} \
|
30 |
+
# --bert_attention_plugin ${INFERENCE_PRECISION} \
|
31 |
+
# --gpt_attention_plugin ${INFERENCE_PRECISION} \
|
32 |
+
# --remove_input_padding disable
|
33 |
+
|
34 |
+
python3 run.py --engine_dir $output_dir --dataset hf-internal-testing/librispeech_asr_dummy --name librispeech_dummy_${output_dir}
|
build_whisper_int8.sh
ADDED
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
export CUDA_VISIBLE_DEVICES=1
|
2 |
+
INFERENCE_PRECISION=float16
|
3 |
+
WEIGHT_ONLY_PRECISION=int8
|
4 |
+
MAX_BEAM_WIDTH=4
|
5 |
+
MAX_BATCH_SIZE=8
|
6 |
+
checkpoint_dir=tllm_checkpoint_${WEIGHT_ONLY_PRECISION}
|
7 |
+
output_dir=distil_whisper_large_v3_${WEIGHT_ONLY_PRECISION}
|
8 |
+
|
9 |
+
# python3 convert_checkpoint.py --use_weight_only \
|
10 |
+
# --weight_only_precision $WEIGHT_ONLY_PRECISION \
|
11 |
+
# --output_dir $checkpoint_dir
|
12 |
+
|
13 |
+
trtllm-build --checkpoint_dir ${checkpoint_dir}/encoder \
|
14 |
+
--output_dir ${output_dir}/encoder \
|
15 |
+
--paged_kv_cache disable \
|
16 |
+
--moe_plugin disable \
|
17 |
+
--enable_xqa disable \
|
18 |
+
--max_batch_size ${MAX_BATCH_SIZE} \
|
19 |
+
--gemm_plugin ${INFERENCE_PRECISION} \
|
20 |
+
--bert_attention_plugin ${INFERENCE_PRECISION} \
|
21 |
+
--remove_input_padding disable
|
22 |
+
|
23 |
+
trtllm-build --checkpoint_dir ${checkpoint_dir}/decoder \
|
24 |
+
--output_dir ${output_dir}/decoder \
|
25 |
+
--paged_kv_cache disable \
|
26 |
+
--moe_plugin disable \
|
27 |
+
--enable_xqa disable \
|
28 |
+
--max_beam_width ${MAX_BEAM_WIDTH} \
|
29 |
+
--max_batch_size ${MAX_BATCH_SIZE} \
|
30 |
+
--max_seq_len 100 \
|
31 |
+
--max_input_len 14 \
|
32 |
+
--max_encoder_input_len 1500 \
|
33 |
+
--gemm_plugin ${INFERENCE_PRECISION} \
|
34 |
+
--bert_attention_plugin ${INFERENCE_PRECISION} \
|
35 |
+
--gpt_attention_plugin ${INFERENCE_PRECISION} \
|
36 |
+
--remove_input_padding disable
|
37 |
+
python3 run.py --engine_dir $output_dir --dataset hf-internal-testing/librispeech_asr_dummy --name librispeech_dummy_${output_dir}
|
tllm_checkpoint/decoder/config.json
ADDED
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architecture": "DecoderModel",
|
3 |
+
"dtype": "float16",
|
4 |
+
"logits_dtype": "float16",
|
5 |
+
"num_hidden_layers": 2,
|
6 |
+
"num_attention_heads": 20,
|
7 |
+
"hidden_size": 1280,
|
8 |
+
"norm_epsilon": 1e-05,
|
9 |
+
"vocab_size": 51866,
|
10 |
+
"hidden_act": "gelu",
|
11 |
+
"use_parallel_embedding": false,
|
12 |
+
"embedding_sharding_dim": 0,
|
13 |
+
"max_position_embeddings": 448,
|
14 |
+
"use_prompt_tuning": false,
|
15 |
+
"head_size": 64,
|
16 |
+
"has_position_embedding": true,
|
17 |
+
"layernorm_type": 0,
|
18 |
+
"has_attention_qkvo_bias": true,
|
19 |
+
"has_mlp_bias": true,
|
20 |
+
"has_model_final_layernorm": true,
|
21 |
+
"has_embedding_layernorm": false,
|
22 |
+
"has_embedding_scale": false,
|
23 |
+
"ffn_hidden_size": 5120,
|
24 |
+
"q_scaling": 1.0,
|
25 |
+
"layernorm_position": 0,
|
26 |
+
"relative_attention": false,
|
27 |
+
"max_distance": 0,
|
28 |
+
"num_buckets": 0,
|
29 |
+
"model_type": "whisper",
|
30 |
+
"rescale_before_lm_head": false,
|
31 |
+
"encoder_hidden_size": 1280,
|
32 |
+
"encoder_num_heads": 20,
|
33 |
+
"encoder_head_size": null,
|
34 |
+
"skip_cross_qkv": false,
|
35 |
+
"quantization": {
|
36 |
+
"quant_algo": null
|
37 |
+
}
|
38 |
+
}
|
tllm_checkpoint/decoder/rank0.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:1a2e1b9d985e4764effd3367fb7994c5b5cc4ac1dedd63f3c829d05f30113118
|
3 |
+
size 371665280
|
tllm_checkpoint/encoder/config.json
ADDED
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architecture": "WhisperEncoder",
|
3 |
+
"dtype": "float16",
|
4 |
+
"num_hidden_layers": 32,
|
5 |
+
"num_attention_heads": 20,
|
6 |
+
"hidden_size": 1280,
|
7 |
+
"n_mels": 128,
|
8 |
+
"n_audio_ctx": 1500,
|
9 |
+
"vocab_size": 51866,
|
10 |
+
"hidden_act": "gelu",
|
11 |
+
"num_languages": 100,
|
12 |
+
"quantization": {
|
13 |
+
"quant_algo": null
|
14 |
+
}
|
15 |
+
}
|
tllm_checkpoint/encoder/rank0.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:bc9f10fdcb40284710f7488bba8c1a021c0d3099487ee886c95c537aacbc22b9
|
3 |
+
size 1288720864
|
tllm_checkpoint_int8/decoder/config.json
ADDED
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architecture": "DecoderModel",
|
3 |
+
"dtype": "float16",
|
4 |
+
"logits_dtype": "float16",
|
5 |
+
"num_hidden_layers": 2,
|
6 |
+
"num_attention_heads": 20,
|
7 |
+
"hidden_size": 1280,
|
8 |
+
"norm_epsilon": 1e-05,
|
9 |
+
"vocab_size": 51866,
|
10 |
+
"hidden_act": "gelu",
|
11 |
+
"use_parallel_embedding": false,
|
12 |
+
"embedding_sharding_dim": 0,
|
13 |
+
"max_position_embeddings": 448,
|
14 |
+
"use_prompt_tuning": false,
|
15 |
+
"head_size": 64,
|
16 |
+
"has_position_embedding": true,
|
17 |
+
"layernorm_type": 0,
|
18 |
+
"has_attention_qkvo_bias": true,
|
19 |
+
"has_mlp_bias": true,
|
20 |
+
"has_model_final_layernorm": true,
|
21 |
+
"has_embedding_layernorm": false,
|
22 |
+
"has_embedding_scale": false,
|
23 |
+
"ffn_hidden_size": 5120,
|
24 |
+
"q_scaling": 1.0,
|
25 |
+
"layernorm_position": 0,
|
26 |
+
"relative_attention": false,
|
27 |
+
"max_distance": 0,
|
28 |
+
"num_buckets": 0,
|
29 |
+
"model_type": "whisper",
|
30 |
+
"rescale_before_lm_head": false,
|
31 |
+
"encoder_hidden_size": 1280,
|
32 |
+
"encoder_num_heads": 20,
|
33 |
+
"encoder_head_size": null,
|
34 |
+
"skip_cross_qkv": false,
|
35 |
+
"quantization": {
|
36 |
+
"quant_algo": "W8A16"
|
37 |
+
}
|
38 |
+
}
|
tllm_checkpoint_int8/decoder/rank0.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:85c059ed4fa6c74025538b2c38b42bce19755dfad6390ee1930dfc5c111fed19
|
3 |
+
size 319304296
|
tllm_checkpoint_int8/encoder/config.json
ADDED
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architecture": "WhisperEncoder",
|
3 |
+
"dtype": "float16",
|
4 |
+
"num_hidden_layers": 32,
|
5 |
+
"num_attention_heads": 20,
|
6 |
+
"hidden_size": 1280,
|
7 |
+
"n_mels": 128,
|
8 |
+
"n_audio_ctx": 1500,
|
9 |
+
"vocab_size": 51866,
|
10 |
+
"hidden_act": "gelu",
|
11 |
+
"num_languages": 100,
|
12 |
+
"quantization": {
|
13 |
+
"quant_algo": "W8A16"
|
14 |
+
}
|
15 |
+
}
|
tllm_checkpoint_int8/encoder/rank0.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d5637200b3af4798f5d775e9c150efbda0084902937aa3a2cf200f9a4dfecb35
|
3 |
+
size 660326480
|