Ship one int8 (8da8w) variant per size
#1
by msluszniak - opened
- .gitattributes +3 -0
- 135m/xnnpack/config.json +4 -52
- 135m/xnnpack/{smollm2_135m_xnnpack_8da4w.pte → smollm2_135m_xnnpack_8da8w.pte} +2 -2
- 1_7b/xnnpack/config.json +4 -52
- 1_7b/xnnpack/{smollm2_1_7b_xnnpack_8da4w.pte → smollm2_1_7b_xnnpack_8da8w.pte} +2 -2
- 1_7b/xnnpack/smollm2_1_7b_xnnpack_bf16.pte +0 -3
- 360m/xnnpack/config.json +4 -52
- 360m/xnnpack/smollm2_360m_xnnpack_8da4w.pte +0 -3
- 135m/xnnpack/smollm2_135m_xnnpack_bf16.pte → 360m/xnnpack/smollm2_360m_xnnpack_8da8w.pte +2 -2
- 360m/xnnpack/smollm2_360m_xnnpack_bf16.pte +0 -3
.gitattributes
CHANGED
|
@@ -46,3 +46,6 @@ smolLm-2-1.7B/quantized/smolLm2_1_7B_8da4w.pte filter=lfs diff=lfs merge=lfs -te
|
|
| 46 |
135m/xnnpack/smollm2_135m_xnnpack_8da4w.pte filter=lfs diff=lfs merge=lfs -text
|
| 47 |
360m/xnnpack/smollm2_360m_xnnpack_bf16.pte filter=lfs diff=lfs merge=lfs -text
|
| 48 |
360m/xnnpack/smollm2_360m_xnnpack_8da4w.pte filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
135m/xnnpack/smollm2_135m_xnnpack_8da4w.pte filter=lfs diff=lfs merge=lfs -text
|
| 47 |
360m/xnnpack/smollm2_360m_xnnpack_bf16.pte filter=lfs diff=lfs merge=lfs -text
|
| 48 |
360m/xnnpack/smollm2_360m_xnnpack_8da4w.pte filter=lfs diff=lfs merge=lfs -text
|
| 49 |
+
135m/xnnpack/smollm2_135m_xnnpack_8da8w.pte filter=lfs diff=lfs merge=lfs -text
|
| 50 |
+
360m/xnnpack/smollm2_360m_xnnpack_8da8w.pte filter=lfs diff=lfs merge=lfs -text
|
| 51 |
+
1_7b/xnnpack/smollm2_1_7b_xnnpack_8da8w.pte filter=lfs diff=lfs merge=lfs -text
|
135m/xnnpack/config.json
CHANGED
|
@@ -9,56 +9,8 @@
|
|
| 9 |
"license": "apache-2.0",
|
| 10 |
"variants": [
|
| 11 |
{
|
| 12 |
-
"file": "
|
| 13 |
-
"precision": "
|
| 14 |
-
"quantized": false,
|
| 15 |
-
"default": true,
|
| 16 |
-
"methods": {
|
| 17 |
-
"enable_dynamic_shape": true,
|
| 18 |
-
"forward": {
|
| 19 |
-
"inputs": [
|
| 20 |
-
{
|
| 21 |
-
"shape": [
|
| 22 |
-
1,
|
| 23 |
-
{
|
| 24 |
-
"min": 1,
|
| 25 |
-
"max": 2048
|
| 26 |
-
}
|
| 27 |
-
],
|
| 28 |
-
"dtype": "int64"
|
| 29 |
-
},
|
| 30 |
-
{
|
| 31 |
-
"shape": [
|
| 32 |
-
1
|
| 33 |
-
],
|
| 34 |
-
"dtype": "int64"
|
| 35 |
-
}
|
| 36 |
-
],
|
| 37 |
-
"outputs": [
|
| 38 |
-
{
|
| 39 |
-
"shape": [
|
| 40 |
-
1,
|
| 41 |
-
49152
|
| 42 |
-
],
|
| 43 |
-
"dtype": "bfloat16"
|
| 44 |
-
}
|
| 45 |
-
]
|
| 46 |
-
},
|
| 47 |
-
"get_bos_id": 1,
|
| 48 |
-
"get_eos_ids": [
|
| 49 |
-
2
|
| 50 |
-
],
|
| 51 |
-
"get_max_context_len": 2048,
|
| 52 |
-
"get_max_seq_len": 2048,
|
| 53 |
-
"get_n_layers": null,
|
| 54 |
-
"get_vocab_size": 49152,
|
| 55 |
-
"use_kv_cache": true,
|
| 56 |
-
"use_sdpa_with_kv_cache": true
|
| 57 |
-
}
|
| 58 |
-
},
|
| 59 |
-
{
|
| 60 |
-
"file": "smollm2_135m_xnnpack_8da4w.pte",
|
| 61 |
-
"precision": "8da4w",
|
| 62 |
"quantized": true,
|
| 63 |
"default": true,
|
| 64 |
"methods": {
|
|
@@ -70,7 +22,7 @@
|
|
| 70 |
1,
|
| 71 |
{
|
| 72 |
"min": 1,
|
| 73 |
-
"max":
|
| 74 |
}
|
| 75 |
],
|
| 76 |
"dtype": "int64"
|
|
@@ -98,7 +50,7 @@
|
|
| 98 |
],
|
| 99 |
"get_max_context_len": 2048,
|
| 100 |
"get_max_seq_len": 2048,
|
| 101 |
-
"get_n_layers":
|
| 102 |
"get_vocab_size": 49152,
|
| 103 |
"use_kv_cache": true,
|
| 104 |
"use_sdpa_with_kv_cache": true
|
|
|
|
| 9 |
"license": "apache-2.0",
|
| 10 |
"variants": [
|
| 11 |
{
|
| 12 |
+
"file": "smollm2_135m_xnnpack_8da8w.pte",
|
| 13 |
+
"precision": "8da8w",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
"quantized": true,
|
| 15 |
"default": true,
|
| 16 |
"methods": {
|
|
|
|
| 22 |
1,
|
| 23 |
{
|
| 24 |
"min": 1,
|
| 25 |
+
"max": 2047
|
| 26 |
}
|
| 27 |
],
|
| 28 |
"dtype": "int64"
|
|
|
|
| 50 |
],
|
| 51 |
"get_max_context_len": 2048,
|
| 52 |
"get_max_seq_len": 2048,
|
| 53 |
+
"get_n_layers": 30,
|
| 54 |
"get_vocab_size": 49152,
|
| 55 |
"use_kv_cache": true,
|
| 56 |
"use_sdpa_with_kv_cache": true
|
135m/xnnpack/{smollm2_135m_xnnpack_8da4w.pte → smollm2_135m_xnnpack_8da8w.pte}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1d95ed83eae1af1c5a8fad5ec5a8893f0270a144a6c087c2c895b357da980eb4
|
| 3 |
+
size 165664000
|
1_7b/xnnpack/config.json
CHANGED
|
@@ -9,56 +9,8 @@
|
|
| 9 |
"license": "apache-2.0",
|
| 10 |
"variants": [
|
| 11 |
{
|
| 12 |
-
"file": "
|
| 13 |
-
"precision": "
|
| 14 |
-
"quantized": false,
|
| 15 |
-
"default": true,
|
| 16 |
-
"methods": {
|
| 17 |
-
"enable_dynamic_shape": true,
|
| 18 |
-
"forward": {
|
| 19 |
-
"inputs": [
|
| 20 |
-
{
|
| 21 |
-
"shape": [
|
| 22 |
-
1,
|
| 23 |
-
{
|
| 24 |
-
"min": 1,
|
| 25 |
-
"max": 2048
|
| 26 |
-
}
|
| 27 |
-
],
|
| 28 |
-
"dtype": "int64"
|
| 29 |
-
},
|
| 30 |
-
{
|
| 31 |
-
"shape": [
|
| 32 |
-
1
|
| 33 |
-
],
|
| 34 |
-
"dtype": "int64"
|
| 35 |
-
}
|
| 36 |
-
],
|
| 37 |
-
"outputs": [
|
| 38 |
-
{
|
| 39 |
-
"shape": [
|
| 40 |
-
1,
|
| 41 |
-
49152
|
| 42 |
-
],
|
| 43 |
-
"dtype": "bfloat16"
|
| 44 |
-
}
|
| 45 |
-
]
|
| 46 |
-
},
|
| 47 |
-
"get_bos_id": 1,
|
| 48 |
-
"get_eos_ids": [
|
| 49 |
-
2
|
| 50 |
-
],
|
| 51 |
-
"get_max_context_len": 2048,
|
| 52 |
-
"get_max_seq_len": 2048,
|
| 53 |
-
"get_n_layers": null,
|
| 54 |
-
"get_vocab_size": 49152,
|
| 55 |
-
"use_kv_cache": true,
|
| 56 |
-
"use_sdpa_with_kv_cache": true
|
| 57 |
-
}
|
| 58 |
-
},
|
| 59 |
-
{
|
| 60 |
-
"file": "smollm2_1_7b_xnnpack_8da4w.pte",
|
| 61 |
-
"precision": "8da4w",
|
| 62 |
"quantized": true,
|
| 63 |
"default": true,
|
| 64 |
"methods": {
|
|
@@ -70,7 +22,7 @@
|
|
| 70 |
1,
|
| 71 |
{
|
| 72 |
"min": 1,
|
| 73 |
-
"max":
|
| 74 |
}
|
| 75 |
],
|
| 76 |
"dtype": "int64"
|
|
@@ -98,7 +50,7 @@
|
|
| 98 |
],
|
| 99 |
"get_max_context_len": 2048,
|
| 100 |
"get_max_seq_len": 2048,
|
| 101 |
-
"get_n_layers":
|
| 102 |
"get_vocab_size": 49152,
|
| 103 |
"use_kv_cache": true,
|
| 104 |
"use_sdpa_with_kv_cache": true
|
|
|
|
| 9 |
"license": "apache-2.0",
|
| 10 |
"variants": [
|
| 11 |
{
|
| 12 |
+
"file": "smollm2_1_7b_xnnpack_8da8w.pte",
|
| 13 |
+
"precision": "8da8w",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
"quantized": true,
|
| 15 |
"default": true,
|
| 16 |
"methods": {
|
|
|
|
| 22 |
1,
|
| 23 |
{
|
| 24 |
"min": 1,
|
| 25 |
+
"max": 2047
|
| 26 |
}
|
| 27 |
],
|
| 28 |
"dtype": "int64"
|
|
|
|
| 50 |
],
|
| 51 |
"get_max_context_len": 2048,
|
| 52 |
"get_max_seq_len": 2048,
|
| 53 |
+
"get_n_layers": 24,
|
| 54 |
"get_vocab_size": 49152,
|
| 55 |
"use_kv_cache": true,
|
| 56 |
"use_sdpa_with_kv_cache": true
|
1_7b/xnnpack/{smollm2_1_7b_xnnpack_8da4w.pte → smollm2_1_7b_xnnpack_8da8w.pte}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:93e86987edccb017b5af3a514be6374d744c2363902f728b837b91b968ca4c66
|
| 3 |
+
size 1816874880
|
1_7b/xnnpack/smollm2_1_7b_xnnpack_bf16.pte
DELETED
|
@@ -1,3 +0,0 @@
|
|
| 1 |
-
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:1540a6f8b48a39819ce384c972c617a619ea9f9ad545cddedff675d9c3926d5c
|
| 3 |
-
size 3423773920
|
|
|
|
|
|
|
|
|
|
|
|
360m/xnnpack/config.json
CHANGED
|
@@ -9,56 +9,8 @@
|
|
| 9 |
"license": "apache-2.0",
|
| 10 |
"variants": [
|
| 11 |
{
|
| 12 |
-
"file": "
|
| 13 |
-
"precision": "
|
| 14 |
-
"quantized": false,
|
| 15 |
-
"default": true,
|
| 16 |
-
"methods": {
|
| 17 |
-
"enable_dynamic_shape": true,
|
| 18 |
-
"forward": {
|
| 19 |
-
"inputs": [
|
| 20 |
-
{
|
| 21 |
-
"shape": [
|
| 22 |
-
1,
|
| 23 |
-
{
|
| 24 |
-
"min": 1,
|
| 25 |
-
"max": 2048
|
| 26 |
-
}
|
| 27 |
-
],
|
| 28 |
-
"dtype": "int64"
|
| 29 |
-
},
|
| 30 |
-
{
|
| 31 |
-
"shape": [
|
| 32 |
-
1
|
| 33 |
-
],
|
| 34 |
-
"dtype": "int64"
|
| 35 |
-
}
|
| 36 |
-
],
|
| 37 |
-
"outputs": [
|
| 38 |
-
{
|
| 39 |
-
"shape": [
|
| 40 |
-
1,
|
| 41 |
-
49152
|
| 42 |
-
],
|
| 43 |
-
"dtype": "bfloat16"
|
| 44 |
-
}
|
| 45 |
-
]
|
| 46 |
-
},
|
| 47 |
-
"get_bos_id": 1,
|
| 48 |
-
"get_eos_ids": [
|
| 49 |
-
2
|
| 50 |
-
],
|
| 51 |
-
"get_max_context_len": 2048,
|
| 52 |
-
"get_max_seq_len": 2048,
|
| 53 |
-
"get_n_layers": null,
|
| 54 |
-
"get_vocab_size": 49152,
|
| 55 |
-
"use_kv_cache": true,
|
| 56 |
-
"use_sdpa_with_kv_cache": true
|
| 57 |
-
}
|
| 58 |
-
},
|
| 59 |
-
{
|
| 60 |
-
"file": "smollm2_360m_xnnpack_8da4w.pte",
|
| 61 |
-
"precision": "8da4w",
|
| 62 |
"quantized": true,
|
| 63 |
"default": true,
|
| 64 |
"methods": {
|
|
@@ -70,7 +22,7 @@
|
|
| 70 |
1,
|
| 71 |
{
|
| 72 |
"min": 1,
|
| 73 |
-
"max":
|
| 74 |
}
|
| 75 |
],
|
| 76 |
"dtype": "int64"
|
|
@@ -98,7 +50,7 @@
|
|
| 98 |
],
|
| 99 |
"get_max_context_len": 2048,
|
| 100 |
"get_max_seq_len": 2048,
|
| 101 |
-
"get_n_layers":
|
| 102 |
"get_vocab_size": 49152,
|
| 103 |
"use_kv_cache": true,
|
| 104 |
"use_sdpa_with_kv_cache": true
|
|
|
|
| 9 |
"license": "apache-2.0",
|
| 10 |
"variants": [
|
| 11 |
{
|
| 12 |
+
"file": "smollm2_360m_xnnpack_8da8w.pte",
|
| 13 |
+
"precision": "8da8w",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
"quantized": true,
|
| 15 |
"default": true,
|
| 16 |
"methods": {
|
|
|
|
| 22 |
1,
|
| 23 |
{
|
| 24 |
"min": 1,
|
| 25 |
+
"max": 2047
|
| 26 |
}
|
| 27 |
],
|
| 28 |
"dtype": "int64"
|
|
|
|
| 50 |
],
|
| 51 |
"get_max_context_len": 2048,
|
| 52 |
"get_max_seq_len": 2048,
|
| 53 |
+
"get_n_layers": 32,
|
| 54 |
"get_vocab_size": 49152,
|
| 55 |
"use_kv_cache": true,
|
| 56 |
"use_sdpa_with_kv_cache": true
|
360m/xnnpack/smollm2_360m_xnnpack_8da4w.pte
DELETED
|
@@ -1,3 +0,0 @@
|
|
| 1 |
-
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:1257ee0804dc3d1ebcfd31ef61979b0a99ae09590bb4d5f3ddd5c85ca419d780
|
| 3 |
-
size 1363730688
|
|
|
|
|
|
|
|
|
|
|
|
135m/xnnpack/smollm2_135m_xnnpack_bf16.pte → 360m/xnnpack/smollm2_360m_xnnpack_8da8w.pte
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d40efc9f3d12547af0d5297aff99c0ee13c8850070b68ef2f93d57154638aab9
|
| 3 |
+
size 412454016
|
360m/xnnpack/smollm2_360m_xnnpack_bf16.pte
DELETED
|
@@ -1,3 +0,0 @@
|
|
| 1 |
-
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:3b88c825a47c75e91e0e18a579188dedd27c5aff312abf3689c58ac1b319c07c
|
| 3 |
-
size 724826208
|
|
|
|
|
|
|
|
|
|
|
|