Upload pplx-embed-v1-4b MLX embedding artifact
Browse files- README.md +47 -150
- config.json +82 -77
- conversion.json +8 -0
- model-00001-of-00002.safetensors +2 -2
- model-00002-of-00002.safetensors +2 -2
- pplx_mlx_convert/__init__.py +14 -0
- pplx_mlx_convert/architecture.py +46 -0
- pplx_mlx_convert/embeddings.py +367 -0
- pplx_mlx_convert/models.py +66 -0
- tokenizer.json +2 -2
- tokenizer_config.json +2 -235
README.md
CHANGED
|
@@ -1,192 +1,89 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
|
|
|
|
|
|
| 3 |
pipeline_tag: feature-extraction
|
| 4 |
tags:
|
|
|
|
|
|
|
| 5 |
- feature-extraction
|
| 6 |
- sentence-similarity
|
| 7 |
- mteb
|
| 8 |
-
-
|
| 9 |
-
|
| 10 |
-
- multilingual
|
| 11 |
---
|
| 12 |
|
|
|
|
| 13 |
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
</p>
|
| 17 |
|
| 18 |
-
|
| 19 |
|
| 20 |
-
|
| 21 |
|
| 22 |
-
|
| 23 |
-
|
|
|
|
| 24 |
|
| 25 |
-
|
| 26 |
-
> `pplx-embed-v1` and `pplx-embed-context-v1` natively produce *unnormalized* int8-quantized embeddings. Ensure that you compare them via *cosine similarity*.
|
| 27 |
|
|
|
|
| 28 |
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
## Models
|
| 32 |
-
|
| 33 |
-
| Model | Dimensions | Context | MRL | Quantization | Instruction | Pooling |
|
| 34 |
-
|:-----:|:----------:|:-------:|:---:|:------------:|:-----------:|:-------:|
|
| 35 |
-
| `pplx-embed-v1-0.6B` | 1024 | 32K | Yes | INT8/BINARY | No | Mean |
|
| 36 |
-
| `pplx-embed-v1-4B` | 2560 | 32K | Yes | INT8/BINARY | No | Mean |
|
| 37 |
-
| `pplx-embed-context-v1-0.6B` | 1024 | 32K | Yes | INT8/BINARY | No | Mean |
|
| 38 |
-
| `pplx-embed-context-v1-4B` | 2560 | 32K | Yes | INT8/BINARY | No | Mean |
|
| 39 |
-
|
| 40 |
-
<sub>All models are built on diffusion continued pre-trained Qwen3 at Perplexity AI.</sub>
|
| 41 |
-
|
| 42 |
-
<sub>Many modern embedding models rely on instruction tuning, where users prepend an instruction string to the text being embedded. This can yield a 2%-3% lift on benchmarks, but it also introduces prompt-selection overhead and can make indexing pipelines brittle (small instruction changes can shift embedding space). We deliberately **avoid** this requirement: you can embed the text you want to index directly, without having to choose or maintain an instruction prefix.</sub>
|
| 43 |
-
|
| 44 |
-
## Usage
|
| 45 |
-
|
| 46 |
-
<details>
|
| 47 |
-
<summary>Via API</summary>
|
| 48 |
|
| 49 |
```bash
|
| 50 |
-
|
| 51 |
-
-H "Authorization: Bearer YOUR_API_KEY" \
|
| 52 |
-
-H "Content-Type: application/json" \
|
| 53 |
-
-d '{
|
| 54 |
-
"input": [
|
| 55 |
-
"Scientists explore the universe driven by curiosity.",
|
| 56 |
-
"Children learn through curious exploration.",
|
| 57 |
-
"Historical discoveries began with curious questions.",
|
| 58 |
-
"Animals use curiosity to adapt and survive.",
|
| 59 |
-
"Philosophy examines the nature of curiosity."
|
| 60 |
-
],
|
| 61 |
-
"model": "pplx-embed-v1-4b"
|
| 62 |
-
}'
|
| 63 |
-
```
|
| 64 |
-
|
| 65 |
-
</details>
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
<details>
|
| 69 |
-
<summary>Using SentenceTransformers</summary>
|
| 70 |
-
|
| 71 |
-
```python
|
| 72 |
-
from sentence_transformers import SentenceTransformer
|
| 73 |
-
|
| 74 |
-
model = SentenceTransformer(
|
| 75 |
-
"perplexity-ai/pplx-embed-v1-4B",
|
| 76 |
-
trust_remote_code=True
|
| 77 |
-
)
|
| 78 |
-
|
| 79 |
-
texts = [
|
| 80 |
-
"Scientists explore the universe driven by curiosity.",
|
| 81 |
-
"Children learn through curious exploration.",
|
| 82 |
-
"Historical discoveries began with curious questions.",
|
| 83 |
-
"Animals use curiosity to adapt and survive.",
|
| 84 |
-
"Philosophy examines the nature of curiosity.",
|
| 85 |
-
]
|
| 86 |
-
|
| 87 |
-
embeddings = model.encode(texts) # Shape: (5, 2560), quantized to int8
|
| 88 |
-
embeddings = model.encode(texts, quantization="binary") # Shape: (5, 2560), quantized to binary
|
| 89 |
```
|
| 90 |
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
<details>
|
| 94 |
-
<summary> Using ONNX models </summary>
|
| 95 |
|
| 96 |
```python
|
|
|
|
|
|
|
| 97 |
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
import numpy as np
|
| 101 |
-
|
| 102 |
-
tokenizer = AutoTokenizer.from_pretrained("perplexity-ai/pplx-embed-v1-4b", trust_remote_code=True)
|
| 103 |
-
session = ort.InferenceSession("onnx/model.onnx")
|
| 104 |
|
|
|
|
| 105 |
|
|
|
|
| 106 |
texts = [
|
| 107 |
"Scientists explore the universe driven by curiosity.",
|
| 108 |
"Children learn through curious exploration.",
|
| 109 |
"Historical discoveries began with curious questions.",
|
| 110 |
-
"Animals use curiosity to adapt and survive.",
|
| 111 |
-
"Philosophy examines the nature of curiosity.",
|
| 112 |
]
|
| 113 |
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
truncation=True,
|
| 118 |
-
return_tensors="np"
|
| 119 |
-
)
|
| 120 |
-
|
| 121 |
-
onnx_inputs = {
|
| 122 |
-
"input_ids": tokenized["input_ids"].astype(np.int64),
|
| 123 |
-
"attention_mask": tokenized["attention_mask"].astype(np.int64),
|
| 124 |
-
}
|
| 125 |
-
|
| 126 |
-
# Run inference
|
| 127 |
-
onnx_embeddings = session.run([out.name for out in session.get_outputs()], onnx_inputs)
|
| 128 |
-
|
| 129 |
-
# ONNX produces both int8 and binary precision embeddings:
|
| 130 |
-
int8_embeddings = onnx_embeddings[2]
|
| 131 |
-
binary_embeddings = onnx_embeddings[3]
|
| 132 |
-
packed_embeddings = np.packbits(binary_embeddings != -1, axis=-1)
|
| 133 |
-
```
|
| 134 |
-
|
| 135 |
-
</details>
|
| 136 |
-
|
| 137 |
-
<details>
|
| 138 |
-
<summary>Using Text Embeddings Inference (TEI)</summary>
|
| 139 |
-
|
| 140 |
-
> [!NOTE]
|
| 141 |
-
> Text Embeddings Inference v1.9.2+ is required.
|
| 142 |
-
|
| 143 |
-
> [!IMPORTANT]
|
| 144 |
-
> Currently, only int8-quantized embeddings are available via TEI. Remember to use cosine similarity with unnormalized int8 embeddings.
|
| 145 |
-
|
| 146 |
-
- CPU w/ Candle:
|
| 147 |
-
|
| 148 |
-
```bash
|
| 149 |
-
docker run -p 8080:80 ghcr.io/huggingface/text-embeddings-inference:cpu-1.9 --model-id perplexity-ai/pplx-embed-v1-4B --dtype float32
|
| 150 |
-
```
|
| 151 |
-
|
| 152 |
-
- CPU w/ ORT (ONNX Runtime):
|
| 153 |
-
|
| 154 |
-
```bash
|
| 155 |
-
docker run -p 8080:80 ghcr.io/huggingface/text-embeddings-inference:cpu-1.9 --model-id onnx-community/pplx-embed-v1-4B --dtype float32
|
| 156 |
```
|
| 157 |
|
| 158 |
-
|
|
|
|
|
|
|
| 159 |
|
| 160 |
-
|
| 161 |
-
docker run --gpus all --shm-size 1g -p 8080:80 ghcr.io/huggingface/text-embeddings-inference:cuda-1.9 --model-id perplexity-ai/pplx-embed-v1-4B --dtype float32
|
| 162 |
-
```
|
| 163 |
|
| 164 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 165 |
|
| 166 |
-
|
| 167 |
-
> container instead of the `cuda-1.9`, as that includes the binaries for Turing, Ampere, Hopper and
|
| 168 |
-
> Blackwell, so using a dedicated container will be lighter e.g., `ampere-1.9`.
|
| 169 |
|
| 170 |
-
|
|
|
|
| 171 |
|
| 172 |
-
|
| 173 |
-
curl http://0.0.0.0:8080/embed \
|
| 174 |
-
-H "Content-Type: application/json" \
|
| 175 |
-
-d '{
|
| 176 |
-
"inputs": [
|
| 177 |
-
"Scientists explore the universe driven by curiosity.",
|
| 178 |
-
"Children learn through curious exploration.",
|
| 179 |
-
"Historical discoveries began with curious questions.",
|
| 180 |
-
"Animals use curiosity to adapt and survive.",
|
| 181 |
-
"Philosophy examines the nature of curiosity."
|
| 182 |
-
],
|
| 183 |
-
"normalize": false
|
| 184 |
-
}'
|
| 185 |
-
```
|
| 186 |
|
| 187 |
-
|
|
|
|
| 188 |
|
| 189 |
-
|
|
|
|
| 190 |
|
| 191 |
-
|
| 192 |
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
+
base_model: perplexity-ai/pplx-embed-v1-4b
|
| 4 |
+
library_name: mlx
|
| 5 |
pipeline_tag: feature-extraction
|
| 6 |
tags:
|
| 7 |
+
- mlx
|
| 8 |
+
- apple-silicon
|
| 9 |
- feature-extraction
|
| 10 |
- sentence-similarity
|
| 11 |
- mteb
|
| 12 |
+
- perplexity
|
| 13 |
+
- qwen3
|
|
|
|
| 14 |
---
|
| 15 |
|
| 16 |
+
# pplx-embed-v1-4b-mlx
|
| 17 |
|
| 18 |
+
MLX conversion of [perplexity-ai/pplx-embed-v1-4b](https://huggingface.co/perplexity-ai/pplx-embed-v1-4b)
|
| 19 |
+
for Apple Silicon.
|
|
|
|
| 20 |
|
| 21 |
+
This is a standard embedding model. It takes a list of texts and returns one embedding matrix for the batch.
|
| 22 |
|
| 23 |
+
## Important Loading Note
|
| 24 |
|
| 25 |
+
This artifact is not loadable through vanilla `mlx_lm.load()` because MLX-LM does not
|
| 26 |
+
natively support Perplexity's custom `bidirectional_pplx_qwen3` model type. The repository
|
| 27 |
+
includes a small `pplx_mlx_convert` loader package for this artifact.
|
| 28 |
|
| 29 |
+
## Source Code
|
|
|
|
| 30 |
|
| 31 |
+
Conversion and validation code lives in [https://github.com/thehumanworks/pplx-mlx](https://github.com/thehumanworks/pplx-mlx).
|
| 32 |
|
| 33 |
+
## Install
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
|
| 35 |
```bash
|
| 36 |
+
pip install mlx mlx-lm transformers huggingface_hub numpy
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
```
|
| 38 |
|
| 39 |
+
## Usage
|
|
|
|
|
|
|
|
|
|
| 40 |
|
| 41 |
```python
|
| 42 |
+
import sys
|
| 43 |
+
from huggingface_hub import snapshot_download
|
| 44 |
|
| 45 |
+
repo_path = snapshot_download("agentmish/pplx-embed-v1-4b-mlx")
|
| 46 |
+
sys.path.insert(0, repo_path)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
|
| 48 |
+
from pplx_mlx_convert import load_embedder
|
| 49 |
|
| 50 |
+
embedder = load_embedder(repo_path)
|
| 51 |
texts = [
|
| 52 |
"Scientists explore the universe driven by curiosity.",
|
| 53 |
"Children learn through curious exploration.",
|
| 54 |
"Historical discoveries began with curious questions.",
|
|
|
|
|
|
|
| 55 |
]
|
| 56 |
|
| 57 |
+
embeddings = embedder.encode(texts)
|
| 58 |
+
print(embeddings.shape) # (3, 2560)
|
| 59 |
+
print(embeddings.dtype) # int8
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
```
|
| 61 |
|
| 62 |
+
The model natively produces unnormalized int8 embeddings by default. Use cosine similarity
|
| 63 |
+
for comparison. `embedder.encode(..., quantization="none")` returns float32 pooled embeddings,
|
| 64 |
+
and `embedder.encode(..., quantization="binary")` returns binary tanh embeddings.
|
| 65 |
|
| 66 |
+
## Conversion Details
|
|
|
|
|
|
|
| 67 |
|
| 68 |
+
- Source model: `perplexity-ai/pplx-embed-v1-4b`
|
| 69 |
+
- Source revision: see `conversion.json`
|
| 70 |
+
- Converted dtype: `bfloat16`
|
| 71 |
+
- Embedding dimension: `2560`
|
| 72 |
+
- Output root expected by this workspace: `artifacts/mlx/pplx-embed-v1-4b`
|
| 73 |
|
| 74 |
+
## Validation
|
|
|
|
|
|
|
| 75 |
|
| 76 |
+
Local MLX smoke validation passed with finite raw float embeddings and int8 embedding output
|
| 77 |
+
shapes `[[2, 2560]]`.
|
| 78 |
|
| 79 |
+
Compared against the original Transformers remote-code float32 model on sample text inputs:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 80 |
|
| 81 |
+
- cosine similarities: 0.9998998, 0.9998891, 0.9999021
|
| 82 |
+
- int8 delta: max absolute int8 delta 2; mean absolute int8 delta 0.259
|
| 83 |
|
| 84 |
+
The MLX artifact is bfloat16 while the reference path used float32, so int8 values are not
|
| 85 |
+
expected to be bit-identical.
|
| 86 |
|
| 87 |
+
## License
|
| 88 |
|
| 89 |
+
The source model is MIT licensed. This conversion preserves the MIT license.
|
config.json
CHANGED
|
@@ -1,79 +1,84 @@
|
|
| 1 |
{
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
"
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
"
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
"rope_theta": 1000000,
|
| 66 |
-
"
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
"vocab_size": 151936,
|
| 75 |
-
"attn_implementation": "sdpa",
|
| 76 |
-
"use_bidirectional_attention": true,
|
| 77 |
-
"source_model_type": "bidirectional_pplx_qwen3",
|
| 78 |
-
"model_file": "mlx_pplx_qwen3.py"
|
| 79 |
-
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"PPLXQwen3Model"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"attn_implementation": "sdpa",
|
| 8 |
+
"auto_map": {
|
| 9 |
+
"AutoConfig": "configuration.PPLXQwen3Config",
|
| 10 |
+
"AutoModel": "modeling.PPLXQwen3Model"
|
| 11 |
+
},
|
| 12 |
+
"bos_token_id": 151643,
|
| 13 |
+
"dtype": "float32",
|
| 14 |
+
"eos_token_id": 151643,
|
| 15 |
+
"head_dim": 128,
|
| 16 |
+
"hidden_act": "silu",
|
| 17 |
+
"hidden_size": 2560,
|
| 18 |
+
"initializer_range": 0.02,
|
| 19 |
+
"intermediate_size": 9728,
|
| 20 |
+
"layer_types": [
|
| 21 |
+
"full_attention",
|
| 22 |
+
"full_attention",
|
| 23 |
+
"full_attention",
|
| 24 |
+
"full_attention",
|
| 25 |
+
"full_attention",
|
| 26 |
+
"full_attention",
|
| 27 |
+
"full_attention",
|
| 28 |
+
"full_attention",
|
| 29 |
+
"full_attention",
|
| 30 |
+
"full_attention",
|
| 31 |
+
"full_attention",
|
| 32 |
+
"full_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"full_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"full_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"full_attention",
|
| 39 |
+
"full_attention",
|
| 40 |
+
"full_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"full_attention",
|
| 43 |
+
"full_attention",
|
| 44 |
+
"full_attention",
|
| 45 |
+
"full_attention",
|
| 46 |
+
"full_attention",
|
| 47 |
+
"full_attention",
|
| 48 |
+
"full_attention",
|
| 49 |
+
"full_attention",
|
| 50 |
+
"full_attention",
|
| 51 |
+
"full_attention",
|
| 52 |
+
"full_attention",
|
| 53 |
+
"full_attention",
|
| 54 |
+
"full_attention",
|
| 55 |
+
"full_attention",
|
| 56 |
+
"full_attention"
|
| 57 |
+
],
|
| 58 |
+
"max_position_embeddings": 32768,
|
| 59 |
+
"max_window_layers": 36,
|
| 60 |
+
"mlx_embedding": {
|
| 61 |
+
"source_repo": "perplexity-ai/pplx-embed-v1-4b",
|
| 62 |
+
"source_revision": "2cd0f789519b81eff4c1ea78cbc5d5e4b1224539",
|
| 63 |
+
"converter": "pplx-mlx-convert",
|
| 64 |
+
"dtype": "bfloat16",
|
| 65 |
+
"kind": "independent"
|
| 66 |
+
},
|
| 67 |
+
"model_type": "bidirectional_pplx_qwen3",
|
| 68 |
+
"num_attention_heads": 32,
|
| 69 |
+
"num_hidden_layers": 36,
|
| 70 |
+
"num_key_value_heads": 8,
|
| 71 |
+
"rms_norm_eps": 1e-06,
|
| 72 |
+
"rope_parameters": {
|
| 73 |
+
"rope_theta": 1000000,
|
| 74 |
+
"rope_type": "default"
|
| 75 |
+
},
|
| 76 |
"rope_theta": 1000000,
|
| 77 |
+
"sliding_window": null,
|
| 78 |
+
"tie_word_embeddings": true,
|
| 79 |
+
"transformers_version": "5.0.0.dev0",
|
| 80 |
+
"use_bidirectional_attention": true,
|
| 81 |
+
"use_cache": false,
|
| 82 |
+
"use_sliding_window": false,
|
| 83 |
+
"vocab_size": 151936
|
| 84 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
conversion.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"slug": "pplx-embed-v1-4b",
|
| 3 |
+
"source_repo": "perplexity-ai/pplx-embed-v1-4b",
|
| 4 |
+
"source_revision": "2cd0f789519b81eff4c1ea78cbc5d5e4b1224539",
|
| 5 |
+
"dtype": "bfloat16",
|
| 6 |
+
"artifact_type": "mlx-independent-embedding",
|
| 7 |
+
"kind": "independent"
|
| 8 |
+
}
|
model-00001-of-00002.safetensors
CHANGED
|
@@ -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 sha256:2dc25292c764c45d20eec2abe4c80b9b879425fae0e42a38032dfa3c028eab43
|
| 3 |
+
size 5321132095
|
model-00002-of-00002.safetensors
CHANGED
|
@@ -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 sha256:6ab68bd4b99c7fec64fbc5a162238a432a6e442b093745fcd3385ff70219362c
|
| 3 |
+
size 2723847400
|
pplx_mlx_convert/__init__.py
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Perplexity model conversion helpers for MLX."""
|
| 2 |
+
|
| 3 |
+
from .embeddings import ContextualEmbedder, IndependentEmbedder, load_embedder
|
| 4 |
+
from .models import MODEL_SPECS, ModelKind, ModelSpec, get_model_spec
|
| 5 |
+
|
| 6 |
+
__all__ = [
|
| 7 |
+
"MODEL_SPECS",
|
| 8 |
+
"ContextualEmbedder",
|
| 9 |
+
"IndependentEmbedder",
|
| 10 |
+
"ModelKind",
|
| 11 |
+
"ModelSpec",
|
| 12 |
+
"get_model_spec",
|
| 13 |
+
"load_embedder",
|
| 14 |
+
]
|
pplx_mlx_convert/architecture.py
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Any
|
| 2 |
+
|
| 3 |
+
import mlx.core as mx
|
| 4 |
+
from mlx_lm.models.qwen3 import ModelArgs, Qwen3Model
|
| 5 |
+
|
| 6 |
+
SUPPORTED_MODEL_TYPES = frozenset({"bidirectional_pplx_qwen3", "qwen3"})
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
def make_bidirectional_padding_mask(attention_mask: mx.array | None) -> mx.array | None:
|
| 10 |
+
"""Create a full-attention key padding mask from a tokenizer attention mask."""
|
| 11 |
+
if attention_mask is None:
|
| 12 |
+
return None
|
| 13 |
+
|
| 14 |
+
return attention_mask.astype(mx.bool_)[:, None, None, :]
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
class PPLXQwen3Model(Qwen3Model):
|
| 18 |
+
"""Qwen3 encoder variant used by Perplexity contextual embedding models."""
|
| 19 |
+
|
| 20 |
+
def __call__(
|
| 21 |
+
self,
|
| 22 |
+
inputs: mx.array,
|
| 23 |
+
attention_mask: mx.array | None = None,
|
| 24 |
+
input_embeddings: mx.array | None = None,
|
| 25 |
+
) -> mx.array:
|
| 26 |
+
if input_embeddings is not None:
|
| 27 |
+
hidden_states = input_embeddings
|
| 28 |
+
else:
|
| 29 |
+
hidden_states = self.embed_tokens(inputs)
|
| 30 |
+
|
| 31 |
+
mask = make_bidirectional_padding_mask(attention_mask)
|
| 32 |
+
|
| 33 |
+
for layer in self.layers:
|
| 34 |
+
hidden_states = layer(hidden_states, mask, None)
|
| 35 |
+
|
| 36 |
+
return self.norm(hidden_states)
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def get_pplx_model_classes(config: dict[str, Any]) -> tuple[type[PPLXQwen3Model], type[ModelArgs]]:
|
| 40 |
+
model_type = config.get("model_type")
|
| 41 |
+
if model_type not in SUPPORTED_MODEL_TYPES:
|
| 42 |
+
supported = ", ".join(sorted(SUPPORTED_MODEL_TYPES))
|
| 43 |
+
msg = f"Unsupported Perplexity model type {model_type!r}; expected one of: {supported}."
|
| 44 |
+
raise ValueError(msg)
|
| 45 |
+
|
| 46 |
+
return PPLXQwen3Model, ModelArgs
|
pplx_mlx_convert/embeddings.py
ADDED
|
@@ -0,0 +1,367 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import json
|
| 2 |
+
from collections.abc import Sequence
|
| 3 |
+
from dataclasses import dataclass
|
| 4 |
+
from pathlib import Path
|
| 5 |
+
from typing import Literal
|
| 6 |
+
|
| 7 |
+
import mlx.core as mx
|
| 8 |
+
import numpy as np
|
| 9 |
+
import numpy.typing as npt
|
| 10 |
+
from huggingface_hub import snapshot_download
|
| 11 |
+
from mlx_lm.utils import load_model
|
| 12 |
+
from transformers import PreTrainedTokenizerBase, Qwen2Tokenizer
|
| 13 |
+
|
| 14 |
+
from .architecture import get_pplx_model_classes
|
| 15 |
+
|
| 16 |
+
Quantization = Literal["int8", "binary", "ubinary", "none"]
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def extract_chunk_token_spans(
|
| 20 |
+
*,
|
| 21 |
+
token_ids: npt.NDArray[np.integer],
|
| 22 |
+
attention_mask: npt.NDArray[np.integer],
|
| 23 |
+
sep_token_id: int,
|
| 24 |
+
) -> list[tuple[int, int]]:
|
| 25 |
+
valid_positions = attention_mask.astype(bool)
|
| 26 |
+
sep_positions = np.flatnonzero((token_ids == sep_token_id) & valid_positions)
|
| 27 |
+
last_valid_pos = int(attention_mask.sum())
|
| 28 |
+
|
| 29 |
+
spans: list[tuple[int, int]] = []
|
| 30 |
+
start_pos = 0
|
| 31 |
+
for sep_pos in sep_positions:
|
| 32 |
+
spans.append((start_pos, int(sep_pos)))
|
| 33 |
+
start_pos = int(sep_pos) + 1
|
| 34 |
+
|
| 35 |
+
spans.append((start_pos, last_valid_pos))
|
| 36 |
+
return spans
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def mean_pool(
|
| 40 |
+
token_embeddings: npt.NDArray[np.floating],
|
| 41 |
+
attention_mask: npt.NDArray[np.integer],
|
| 42 |
+
) -> npt.NDArray[np.float32]:
|
| 43 |
+
if token_embeddings.shape[0] == 0:
|
| 44 |
+
return np.zeros(token_embeddings.shape[-1], dtype=np.float32)
|
| 45 |
+
|
| 46 |
+
mask = attention_mask.astype(np.float32)[:, None]
|
| 47 |
+
denominator = np.clip(mask.sum(axis=0), a_min=1e-9, a_max=None)
|
| 48 |
+
return ((token_embeddings * mask).sum(axis=0) / denominator).astype(np.float32)
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
def quantize_int8_tanh(values: npt.NDArray[np.floating]) -> npt.NDArray[np.int8]:
|
| 52 |
+
rounded = np.round(np.tanh(values) * 127)
|
| 53 |
+
return np.clip(rounded, -128, 127).astype(np.int8)
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
def quantize_binary_tanh(values: npt.NDArray[np.floating]) -> npt.NDArray[np.float32]:
|
| 57 |
+
return np.where(values >= 0, 1.0, -1.0).astype(np.float32)
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
def quantize_ubinary_tanh(values: npt.NDArray[np.floating]) -> npt.NDArray[np.uint8]:
|
| 61 |
+
return np.packbits(values >= 0, axis=-1)
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
@dataclass(frozen=True, slots=True)
|
| 65 |
+
class SmokeValidationResult:
|
| 66 |
+
documents: int
|
| 67 |
+
chunk_counts: tuple[int, ...]
|
| 68 |
+
shapes: tuple[tuple[int, ...], ...]
|
| 69 |
+
dtypes: tuple[str, ...]
|
| 70 |
+
raw_float_finite: bool
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
class ContextualEmbedder:
|
| 74 |
+
def __init__(self, model_path: Path | str, *, revision: str | None = None) -> None:
|
| 75 |
+
self.model_path = _resolve_model_path(model_path, revision=revision)
|
| 76 |
+
self.model, self.config = load_model(
|
| 77 |
+
self.model_path,
|
| 78 |
+
lazy=False,
|
| 79 |
+
get_model_classes=get_pplx_model_classes,
|
| 80 |
+
)
|
| 81 |
+
self.tokenizer: PreTrainedTokenizerBase = Qwen2Tokenizer.from_pretrained(self.model_path)
|
| 82 |
+
|
| 83 |
+
if self.tokenizer.sep_token_id is None:
|
| 84 |
+
raise ValueError("Tokenizer must define sep_token_id for contextual chunk extraction.")
|
| 85 |
+
|
| 86 |
+
def encode(
|
| 87 |
+
self,
|
| 88 |
+
documents: Sequence[Sequence[str]],
|
| 89 |
+
*,
|
| 90 |
+
batch_size: int = 4,
|
| 91 |
+
quantization: Quantization = "int8",
|
| 92 |
+
normalize_embeddings: bool = False,
|
| 93 |
+
dimensions: int | None = None,
|
| 94 |
+
) -> list[npt.NDArray[np.generic]]:
|
| 95 |
+
_validate_documents(documents)
|
| 96 |
+
if quantization not in {"int8", "binary", "none"}:
|
| 97 |
+
msg = f"Unsupported quantization {quantization!r}."
|
| 98 |
+
raise ValueError(msg)
|
| 99 |
+
|
| 100 |
+
encoded: list[npt.NDArray[np.generic]] = []
|
| 101 |
+
for start in range(0, len(documents), batch_size):
|
| 102 |
+
batch_docs = documents[start : start + batch_size]
|
| 103 |
+
encoded.extend(
|
| 104 |
+
self._encode_batch(
|
| 105 |
+
batch_docs,
|
| 106 |
+
quantization=quantization,
|
| 107 |
+
normalize_embeddings=normalize_embeddings,
|
| 108 |
+
dimensions=dimensions,
|
| 109 |
+
)
|
| 110 |
+
)
|
| 111 |
+
|
| 112 |
+
return encoded
|
| 113 |
+
|
| 114 |
+
def smoke_validate(self) -> SmokeValidationResult:
|
| 115 |
+
sample_documents = [
|
| 116 |
+
[
|
| 117 |
+
"Curiosity begins in childhood with questions about the world.",
|
| 118 |
+
"Scientific breakthroughs often start with a curious question.",
|
| 119 |
+
],
|
| 120 |
+
[
|
| 121 |
+
"The Mars rover searches for signs of ancient life.",
|
| 122 |
+
],
|
| 123 |
+
]
|
| 124 |
+
raw_embeddings = self.encode(sample_documents, quantization="none")
|
| 125 |
+
raw_float_finite = all(bool(np.isfinite(embedding).all()) for embedding in raw_embeddings)
|
| 126 |
+
embeddings = self.encode(sample_documents, quantization="int8")
|
| 127 |
+
return SmokeValidationResult(
|
| 128 |
+
documents=len(embeddings),
|
| 129 |
+
chunk_counts=tuple(int(embedding.shape[0]) for embedding in embeddings),
|
| 130 |
+
shapes=tuple(tuple(int(dim) for dim in embedding.shape) for embedding in embeddings),
|
| 131 |
+
dtypes=tuple(str(embedding.dtype) for embedding in embeddings),
|
| 132 |
+
raw_float_finite=raw_float_finite,
|
| 133 |
+
)
|
| 134 |
+
|
| 135 |
+
def _encode_batch(
|
| 136 |
+
self,
|
| 137 |
+
documents: Sequence[Sequence[str]],
|
| 138 |
+
*,
|
| 139 |
+
quantization: Quantization,
|
| 140 |
+
normalize_embeddings: bool,
|
| 141 |
+
dimensions: int | None,
|
| 142 |
+
) -> list[npt.NDArray[np.generic]]:
|
| 143 |
+
sep_token = self.tokenizer.sep_token
|
| 144 |
+
if sep_token is None:
|
| 145 |
+
raise ValueError("Tokenizer must define sep_token for contextual chunk joining.")
|
| 146 |
+
|
| 147 |
+
doc_strings = [sep_token.join(chunks) for chunks in documents]
|
| 148 |
+
inputs = self.tokenizer(
|
| 149 |
+
doc_strings,
|
| 150 |
+
padding=True,
|
| 151 |
+
truncation=True,
|
| 152 |
+
return_tensors="np",
|
| 153 |
+
)
|
| 154 |
+
input_ids = np.asarray(inputs["input_ids"])
|
| 155 |
+
attention_mask = np.asarray(inputs["attention_mask"])
|
| 156 |
+
|
| 157 |
+
token_embeddings = self.model(
|
| 158 |
+
mx.array(input_ids),
|
| 159 |
+
attention_mask=mx.array(attention_mask),
|
| 160 |
+
)
|
| 161 |
+
token_embeddings = token_embeddings.astype(mx.float32)
|
| 162 |
+
mx.eval(token_embeddings)
|
| 163 |
+
token_embeddings_np = np.asarray(token_embeddings)
|
| 164 |
+
|
| 165 |
+
batch_embeddings: list[npt.NDArray[np.generic]] = []
|
| 166 |
+
for batch_index in range(len(documents)):
|
| 167 |
+
spans = extract_chunk_token_spans(
|
| 168 |
+
token_ids=input_ids[batch_index],
|
| 169 |
+
attention_mask=attention_mask[batch_index],
|
| 170 |
+
sep_token_id=int(self.tokenizer.sep_token_id),
|
| 171 |
+
)
|
| 172 |
+
chunk_embeddings = [
|
| 173 |
+
mean_pool(
|
| 174 |
+
token_embeddings_np[batch_index, span_start:span_end],
|
| 175 |
+
attention_mask[batch_index, span_start:span_end],
|
| 176 |
+
)
|
| 177 |
+
for span_start, span_end in spans
|
| 178 |
+
]
|
| 179 |
+
stacked = np.stack(chunk_embeddings, axis=0)
|
| 180 |
+
if dimensions is not None:
|
| 181 |
+
stacked = stacked[..., :dimensions]
|
| 182 |
+
batch_embeddings.append(
|
| 183 |
+
_finalize_embeddings(
|
| 184 |
+
stacked,
|
| 185 |
+
quantization=quantization,
|
| 186 |
+
normalize_embeddings=normalize_embeddings,
|
| 187 |
+
)
|
| 188 |
+
)
|
| 189 |
+
|
| 190 |
+
return batch_embeddings
|
| 191 |
+
|
| 192 |
+
|
| 193 |
+
class IndependentEmbedder:
|
| 194 |
+
def __init__(self, model_path: Path | str, *, revision: str | None = None) -> None:
|
| 195 |
+
self.model_path = _resolve_model_path(model_path, revision=revision)
|
| 196 |
+
self.model, self.config = load_model(
|
| 197 |
+
self.model_path,
|
| 198 |
+
lazy=False,
|
| 199 |
+
get_model_classes=get_pplx_model_classes,
|
| 200 |
+
)
|
| 201 |
+
self.tokenizer: PreTrainedTokenizerBase = Qwen2Tokenizer.from_pretrained(self.model_path)
|
| 202 |
+
|
| 203 |
+
def encode(
|
| 204 |
+
self,
|
| 205 |
+
texts: Sequence[str],
|
| 206 |
+
*,
|
| 207 |
+
batch_size: int = 4,
|
| 208 |
+
quantization: Quantization = "int8",
|
| 209 |
+
normalize_embeddings: bool = False,
|
| 210 |
+
dimensions: int | None = None,
|
| 211 |
+
) -> npt.NDArray[np.generic]:
|
| 212 |
+
_validate_texts(texts)
|
| 213 |
+
if quantization not in {"int8", "binary", "ubinary", "none"}:
|
| 214 |
+
msg = f"Unsupported quantization {quantization!r}."
|
| 215 |
+
raise ValueError(msg)
|
| 216 |
+
|
| 217 |
+
encoded_batches: list[npt.NDArray[np.generic]] = []
|
| 218 |
+
for start in range(0, len(texts), batch_size):
|
| 219 |
+
encoded_batches.append(
|
| 220 |
+
self._encode_batch(
|
| 221 |
+
texts[start : start + batch_size],
|
| 222 |
+
quantization=quantization,
|
| 223 |
+
normalize_embeddings=normalize_embeddings,
|
| 224 |
+
dimensions=dimensions,
|
| 225 |
+
)
|
| 226 |
+
)
|
| 227 |
+
return np.concatenate(encoded_batches, axis=0)
|
| 228 |
+
|
| 229 |
+
def smoke_validate(self) -> SmokeValidationResult:
|
| 230 |
+
sample_texts = [
|
| 231 |
+
"Scientists explore the universe driven by curiosity.",
|
| 232 |
+
"Children learn through curious exploration.",
|
| 233 |
+
]
|
| 234 |
+
raw_embeddings = self.encode(sample_texts, quantization="none")
|
| 235 |
+
raw_float_finite = bool(np.isfinite(raw_embeddings).all())
|
| 236 |
+
embeddings = self.encode(sample_texts, quantization="int8")
|
| 237 |
+
return SmokeValidationResult(
|
| 238 |
+
documents=int(embeddings.shape[0]),
|
| 239 |
+
chunk_counts=(),
|
| 240 |
+
shapes=(tuple(int(dim) for dim in embeddings.shape),),
|
| 241 |
+
dtypes=(str(embeddings.dtype),),
|
| 242 |
+
raw_float_finite=raw_float_finite,
|
| 243 |
+
)
|
| 244 |
+
|
| 245 |
+
def _encode_batch(
|
| 246 |
+
self,
|
| 247 |
+
texts: Sequence[str],
|
| 248 |
+
*,
|
| 249 |
+
quantization: Quantization,
|
| 250 |
+
normalize_embeddings: bool,
|
| 251 |
+
dimensions: int | None,
|
| 252 |
+
) -> npt.NDArray[np.generic]:
|
| 253 |
+
inputs = self.tokenizer(
|
| 254 |
+
list(texts),
|
| 255 |
+
padding=True,
|
| 256 |
+
truncation=True,
|
| 257 |
+
return_tensors="np",
|
| 258 |
+
)
|
| 259 |
+
input_ids = np.asarray(inputs["input_ids"])
|
| 260 |
+
attention_mask = np.asarray(inputs["attention_mask"])
|
| 261 |
+
token_embeddings = self.model(
|
| 262 |
+
mx.array(input_ids),
|
| 263 |
+
attention_mask=mx.array(attention_mask),
|
| 264 |
+
)
|
| 265 |
+
token_embeddings = token_embeddings.astype(mx.float32)
|
| 266 |
+
mx.eval(token_embeddings)
|
| 267 |
+
token_embeddings_np = np.asarray(token_embeddings)
|
| 268 |
+
pooled = np.stack(
|
| 269 |
+
[
|
| 270 |
+
mean_pool(token_embeddings_np[row_index], attention_mask[row_index])
|
| 271 |
+
for row_index in range(len(texts))
|
| 272 |
+
],
|
| 273 |
+
axis=0,
|
| 274 |
+
)
|
| 275 |
+
if dimensions is not None:
|
| 276 |
+
pooled = pooled[..., :dimensions]
|
| 277 |
+
return _finalize_embeddings(
|
| 278 |
+
pooled,
|
| 279 |
+
quantization=quantization,
|
| 280 |
+
normalize_embeddings=normalize_embeddings,
|
| 281 |
+
)
|
| 282 |
+
|
| 283 |
+
|
| 284 |
+
Embedder = ContextualEmbedder | IndependentEmbedder
|
| 285 |
+
|
| 286 |
+
|
| 287 |
+
def load_embedder(model_path: Path | str, *, revision: str | None = None) -> Embedder:
|
| 288 |
+
resolved_path = _resolve_model_path(model_path, revision=revision)
|
| 289 |
+
config = json.loads((resolved_path / "config.json").read_text())
|
| 290 |
+
if _is_independent_embedding_config(config):
|
| 291 |
+
return IndependentEmbedder(resolved_path)
|
| 292 |
+
return ContextualEmbedder(resolved_path)
|
| 293 |
+
|
| 294 |
+
|
| 295 |
+
def _is_independent_embedding_config(config: dict[str, object]) -> bool:
|
| 296 |
+
metadata = config.get("mlx_embedding")
|
| 297 |
+
if isinstance(metadata, dict) and metadata.get("kind") == "independent":
|
| 298 |
+
return True
|
| 299 |
+
|
| 300 |
+
auto_map = config.get("auto_map")
|
| 301 |
+
if isinstance(auto_map, dict):
|
| 302 |
+
auto_model = auto_map.get("AutoModel")
|
| 303 |
+
return auto_model == "modeling.PPLXQwen3Model"
|
| 304 |
+
|
| 305 |
+
return False
|
| 306 |
+
|
| 307 |
+
|
| 308 |
+
def _resolve_model_path(model_path: Path | str, *, revision: str | None = None) -> Path:
|
| 309 |
+
path = Path(model_path).expanduser()
|
| 310 |
+
if path.exists():
|
| 311 |
+
return path
|
| 312 |
+
|
| 313 |
+
return Path(
|
| 314 |
+
snapshot_download(
|
| 315 |
+
str(model_path),
|
| 316 |
+
revision=revision,
|
| 317 |
+
allow_patterns=[
|
| 318 |
+
"*.json",
|
| 319 |
+
"*.safetensors",
|
| 320 |
+
"*.py",
|
| 321 |
+
"*.txt",
|
| 322 |
+
"pplx_mlx_convert/*.py",
|
| 323 |
+
],
|
| 324 |
+
)
|
| 325 |
+
)
|
| 326 |
+
|
| 327 |
+
|
| 328 |
+
def _validate_documents(documents: Sequence[Sequence[str]]) -> None:
|
| 329 |
+
if not documents:
|
| 330 |
+
raise ValueError("documents must contain at least one document.")
|
| 331 |
+
for document in documents:
|
| 332 |
+
if not document:
|
| 333 |
+
raise ValueError("Each document must contain at least one chunk.")
|
| 334 |
+
for chunk in document:
|
| 335 |
+
if not isinstance(chunk, str) or not chunk:
|
| 336 |
+
raise ValueError("Each chunk must be a non-empty string.")
|
| 337 |
+
|
| 338 |
+
|
| 339 |
+
def _validate_texts(texts: Sequence[str]) -> None:
|
| 340 |
+
if not texts:
|
| 341 |
+
raise ValueError("texts must contain at least one text.")
|
| 342 |
+
for text in texts:
|
| 343 |
+
if not isinstance(text, str) or not text:
|
| 344 |
+
raise ValueError("Each text must be a non-empty string.")
|
| 345 |
+
|
| 346 |
+
|
| 347 |
+
def _finalize_embeddings(
|
| 348 |
+
embeddings: npt.NDArray[np.floating],
|
| 349 |
+
*,
|
| 350 |
+
quantization: Quantization,
|
| 351 |
+
normalize_embeddings: bool,
|
| 352 |
+
) -> npt.NDArray[np.generic]:
|
| 353 |
+
if quantization == "int8":
|
| 354 |
+
finalized: npt.NDArray[np.generic] = quantize_int8_tanh(embeddings)
|
| 355 |
+
elif quantization == "binary":
|
| 356 |
+
finalized = quantize_binary_tanh(embeddings)
|
| 357 |
+
elif quantization == "ubinary":
|
| 358 |
+
finalized = quantize_ubinary_tanh(embeddings)
|
| 359 |
+
else:
|
| 360 |
+
finalized = embeddings.astype(np.float32)
|
| 361 |
+
|
| 362 |
+
if normalize_embeddings:
|
| 363 |
+
float_embeddings = finalized.astype(np.float32)
|
| 364 |
+
norms = np.linalg.norm(float_embeddings, axis=-1, keepdims=True)
|
| 365 |
+
finalized = float_embeddings / np.clip(norms, a_min=1e-12, a_max=None)
|
| 366 |
+
|
| 367 |
+
return finalized
|
pplx_mlx_convert/models.py
ADDED
|
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from dataclasses import dataclass
|
| 2 |
+
from typing import Final, Literal
|
| 3 |
+
|
| 4 |
+
ModelKind = Literal["independent", "contextual"]
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
@dataclass(frozen=True, slots=True)
|
| 8 |
+
class ModelSpec:
|
| 9 |
+
"""Metadata for one source model that this workspace will convert."""
|
| 10 |
+
|
| 11 |
+
slug: str
|
| 12 |
+
huggingface_repo: str
|
| 13 |
+
parameter_count: str
|
| 14 |
+
embedding_dimension: int
|
| 15 |
+
kind: ModelKind
|
| 16 |
+
|
| 17 |
+
@property
|
| 18 |
+
def huggingface_url(self) -> str:
|
| 19 |
+
return f"https://huggingface.co/{self.huggingface_repo}"
|
| 20 |
+
|
| 21 |
+
@property
|
| 22 |
+
def default_output_dir(self) -> str:
|
| 23 |
+
return f"artifacts/mlx/{self.slug}"
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
MODEL_SPECS: Final[tuple[ModelSpec, ...]] = (
|
| 27 |
+
ModelSpec(
|
| 28 |
+
slug="pplx-embed-v1-4b",
|
| 29 |
+
huggingface_repo="perplexity-ai/pplx-embed-v1-4b",
|
| 30 |
+
parameter_count="4b",
|
| 31 |
+
embedding_dimension=2560,
|
| 32 |
+
kind="independent",
|
| 33 |
+
),
|
| 34 |
+
ModelSpec(
|
| 35 |
+
slug="pplx-embed-v1-0.6b",
|
| 36 |
+
huggingface_repo="perplexity-ai/pplx-embed-v1-0.6b",
|
| 37 |
+
parameter_count="0.6b",
|
| 38 |
+
embedding_dimension=1024,
|
| 39 |
+
kind="independent",
|
| 40 |
+
),
|
| 41 |
+
ModelSpec(
|
| 42 |
+
slug="pplx-embed-context-v1-4b",
|
| 43 |
+
huggingface_repo="perplexity-ai/pplx-embed-context-v1-4b",
|
| 44 |
+
parameter_count="4b",
|
| 45 |
+
embedding_dimension=2560,
|
| 46 |
+
kind="contextual",
|
| 47 |
+
),
|
| 48 |
+
ModelSpec(
|
| 49 |
+
slug="pplx-embed-context-v1-0.6b",
|
| 50 |
+
huggingface_repo="perplexity-ai/pplx-embed-context-v1-0.6b",
|
| 51 |
+
parameter_count="0.6b",
|
| 52 |
+
embedding_dimension=1024,
|
| 53 |
+
kind="contextual",
|
| 54 |
+
),
|
| 55 |
+
)
|
| 56 |
+
|
| 57 |
+
MODEL_SPECS_BY_SLUG: Final[dict[str, ModelSpec]] = {spec.slug: spec for spec in MODEL_SPECS}
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
def get_model_spec(slug: str) -> ModelSpec:
|
| 61 |
+
try:
|
| 62 |
+
return MODEL_SPECS_BY_SLUG[slug]
|
| 63 |
+
except KeyError as exc:
|
| 64 |
+
available = ", ".join(MODEL_SPECS_BY_SLUG)
|
| 65 |
+
msg = f"Unknown model slug {slug!r}. Available slugs: {available}."
|
| 66 |
+
raise ValueError(msg) from exc
|
tokenizer.json
CHANGED
|
@@ -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 sha256:32687b48a8d7da95d23b32a8f24677795496605001bddee04016bb78ebcc2e67
|
| 3 |
+
size 11422833
|
tokenizer_config.json
CHANGED
|
@@ -1,244 +1,11 @@
|
|
| 1 |
{
|
| 2 |
-
"add_bos_token": false,
|
| 3 |
"add_prefix_space": false,
|
| 4 |
-
"
|
| 5 |
-
"151642": {
|
| 6 |
-
"content": "â½Ĺ",
|
| 7 |
-
"lstrip": false,
|
| 8 |
-
"normalized": false,
|
| 9 |
-
"rstrip": false,
|
| 10 |
-
"single_word": false,
|
| 11 |
-
"special": true
|
| 12 |
-
},
|
| 13 |
-
"151643": {
|
| 14 |
-
"content": "<|endoftext|>",
|
| 15 |
-
"lstrip": false,
|
| 16 |
-
"normalized": false,
|
| 17 |
-
"rstrip": false,
|
| 18 |
-
"single_word": false,
|
| 19 |
-
"special": true
|
| 20 |
-
},
|
| 21 |
-
"151644": {
|
| 22 |
-
"content": "<|im_start|>",
|
| 23 |
-
"lstrip": false,
|
| 24 |
-
"normalized": false,
|
| 25 |
-
"rstrip": false,
|
| 26 |
-
"single_word": false,
|
| 27 |
-
"special": true
|
| 28 |
-
},
|
| 29 |
-
"151645": {
|
| 30 |
-
"content": "<|im_end|>",
|
| 31 |
-
"lstrip": false,
|
| 32 |
-
"normalized": false,
|
| 33 |
-
"rstrip": false,
|
| 34 |
-
"single_word": false,
|
| 35 |
-
"special": true
|
| 36 |
-
},
|
| 37 |
-
"151646": {
|
| 38 |
-
"content": "<|object_ref_start|>",
|
| 39 |
-
"lstrip": false,
|
| 40 |
-
"normalized": false,
|
| 41 |
-
"rstrip": false,
|
| 42 |
-
"single_word": false,
|
| 43 |
-
"special": true
|
| 44 |
-
},
|
| 45 |
-
"151647": {
|
| 46 |
-
"content": "<|object_ref_end|>",
|
| 47 |
-
"lstrip": false,
|
| 48 |
-
"normalized": false,
|
| 49 |
-
"rstrip": false,
|
| 50 |
-
"single_word": false,
|
| 51 |
-
"special": true
|
| 52 |
-
},
|
| 53 |
-
"151648": {
|
| 54 |
-
"content": "<|box_start|>",
|
| 55 |
-
"lstrip": false,
|
| 56 |
-
"normalized": false,
|
| 57 |
-
"rstrip": false,
|
| 58 |
-
"single_word": false,
|
| 59 |
-
"special": true
|
| 60 |
-
},
|
| 61 |
-
"151649": {
|
| 62 |
-
"content": "<|box_end|>",
|
| 63 |
-
"lstrip": false,
|
| 64 |
-
"normalized": false,
|
| 65 |
-
"rstrip": false,
|
| 66 |
-
"single_word": false,
|
| 67 |
-
"special": true
|
| 68 |
-
},
|
| 69 |
-
"151650": {
|
| 70 |
-
"content": "<|quad_start|>",
|
| 71 |
-
"lstrip": false,
|
| 72 |
-
"normalized": false,
|
| 73 |
-
"rstrip": false,
|
| 74 |
-
"single_word": false,
|
| 75 |
-
"special": true
|
| 76 |
-
},
|
| 77 |
-
"151651": {
|
| 78 |
-
"content": "<|quad_end|>",
|
| 79 |
-
"lstrip": false,
|
| 80 |
-
"normalized": false,
|
| 81 |
-
"rstrip": false,
|
| 82 |
-
"single_word": false,
|
| 83 |
-
"special": true
|
| 84 |
-
},
|
| 85 |
-
"151652": {
|
| 86 |
-
"content": "<|vision_start|>",
|
| 87 |
-
"lstrip": false,
|
| 88 |
-
"normalized": false,
|
| 89 |
-
"rstrip": false,
|
| 90 |
-
"single_word": false,
|
| 91 |
-
"special": true
|
| 92 |
-
},
|
| 93 |
-
"151653": {
|
| 94 |
-
"content": "<|vision_end|>",
|
| 95 |
-
"lstrip": false,
|
| 96 |
-
"normalized": false,
|
| 97 |
-
"rstrip": false,
|
| 98 |
-
"single_word": false,
|
| 99 |
-
"special": true
|
| 100 |
-
},
|
| 101 |
-
"151654": {
|
| 102 |
-
"content": "<|vision_pad|>",
|
| 103 |
-
"lstrip": false,
|
| 104 |
-
"normalized": false,
|
| 105 |
-
"rstrip": false,
|
| 106 |
-
"single_word": false,
|
| 107 |
-
"special": true
|
| 108 |
-
},
|
| 109 |
-
"151655": {
|
| 110 |
-
"content": "<|image_pad|>",
|
| 111 |
-
"lstrip": false,
|
| 112 |
-
"normalized": false,
|
| 113 |
-
"rstrip": false,
|
| 114 |
-
"single_word": false,
|
| 115 |
-
"special": true
|
| 116 |
-
},
|
| 117 |
-
"151656": {
|
| 118 |
-
"content": "<|video_pad|>",
|
| 119 |
-
"lstrip": false,
|
| 120 |
-
"normalized": false,
|
| 121 |
-
"rstrip": false,
|
| 122 |
-
"single_word": false,
|
| 123 |
-
"special": true
|
| 124 |
-
},
|
| 125 |
-
"151657": {
|
| 126 |
-
"content": "<tool_call>",
|
| 127 |
-
"lstrip": false,
|
| 128 |
-
"normalized": false,
|
| 129 |
-
"rstrip": false,
|
| 130 |
-
"single_word": false,
|
| 131 |
-
"special": false
|
| 132 |
-
},
|
| 133 |
-
"151658": {
|
| 134 |
-
"content": "</tool_call>",
|
| 135 |
-
"lstrip": false,
|
| 136 |
-
"normalized": false,
|
| 137 |
-
"rstrip": false,
|
| 138 |
-
"single_word": false,
|
| 139 |
-
"special": false
|
| 140 |
-
},
|
| 141 |
-
"151659": {
|
| 142 |
-
"content": "<|fim_prefix|>",
|
| 143 |
-
"lstrip": false,
|
| 144 |
-
"normalized": false,
|
| 145 |
-
"rstrip": false,
|
| 146 |
-
"single_word": false,
|
| 147 |
-
"special": false
|
| 148 |
-
},
|
| 149 |
-
"151660": {
|
| 150 |
-
"content": "<|fim_middle|>",
|
| 151 |
-
"lstrip": false,
|
| 152 |
-
"normalized": false,
|
| 153 |
-
"rstrip": false,
|
| 154 |
-
"single_word": false,
|
| 155 |
-
"special": false
|
| 156 |
-
},
|
| 157 |
-
"151661": {
|
| 158 |
-
"content": "<|fim_suffix|>",
|
| 159 |
-
"lstrip": false,
|
| 160 |
-
"normalized": false,
|
| 161 |
-
"rstrip": false,
|
| 162 |
-
"single_word": false,
|
| 163 |
-
"special": false
|
| 164 |
-
},
|
| 165 |
-
"151662": {
|
| 166 |
-
"content": "<|fim_pad|>",
|
| 167 |
-
"lstrip": false,
|
| 168 |
-
"normalized": false,
|
| 169 |
-
"rstrip": false,
|
| 170 |
-
"single_word": false,
|
| 171 |
-
"special": false
|
| 172 |
-
},
|
| 173 |
-
"151663": {
|
| 174 |
-
"content": "<|repo_name|>",
|
| 175 |
-
"lstrip": false,
|
| 176 |
-
"normalized": false,
|
| 177 |
-
"rstrip": false,
|
| 178 |
-
"single_word": false,
|
| 179 |
-
"special": false
|
| 180 |
-
},
|
| 181 |
-
"151664": {
|
| 182 |
-
"content": "<|file_sep|>",
|
| 183 |
-
"lstrip": false,
|
| 184 |
-
"normalized": false,
|
| 185 |
-
"rstrip": false,
|
| 186 |
-
"single_word": false,
|
| 187 |
-
"special": false
|
| 188 |
-
},
|
| 189 |
-
"151665": {
|
| 190 |
-
"content": "<tool_response>",
|
| 191 |
-
"lstrip": false,
|
| 192 |
-
"normalized": false,
|
| 193 |
-
"rstrip": false,
|
| 194 |
-
"single_word": false,
|
| 195 |
-
"special": false
|
| 196 |
-
},
|
| 197 |
-
"151666": {
|
| 198 |
-
"content": "</tool_response>",
|
| 199 |
-
"lstrip": false,
|
| 200 |
-
"normalized": false,
|
| 201 |
-
"rstrip": false,
|
| 202 |
-
"single_word": false,
|
| 203 |
-
"special": false
|
| 204 |
-
},
|
| 205 |
-
"151667": {
|
| 206 |
-
"content": "<think>",
|
| 207 |
-
"lstrip": false,
|
| 208 |
-
"normalized": false,
|
| 209 |
-
"rstrip": false,
|
| 210 |
-
"single_word": false,
|
| 211 |
-
"special": false
|
| 212 |
-
},
|
| 213 |
-
"151668": {
|
| 214 |
-
"content": "</think>",
|
| 215 |
-
"lstrip": false,
|
| 216 |
-
"normalized": false,
|
| 217 |
-
"rstrip": false,
|
| 218 |
-
"single_word": false,
|
| 219 |
-
"special": false
|
| 220 |
-
}
|
| 221 |
-
},
|
| 222 |
-
"additional_special_tokens": [
|
| 223 |
-
"<|im_start|>",
|
| 224 |
-
"<|im_end|>",
|
| 225 |
-
"<|object_ref_start|>",
|
| 226 |
-
"<|object_ref_end|>",
|
| 227 |
-
"<|box_start|>",
|
| 228 |
-
"<|box_end|>",
|
| 229 |
-
"<|quad_start|>",
|
| 230 |
-
"<|quad_end|>",
|
| 231 |
-
"<|vision_start|>",
|
| 232 |
-
"<|vision_end|>",
|
| 233 |
-
"<|vision_pad|>",
|
| 234 |
-
"<|image_pad|>",
|
| 235 |
-
"<|video_pad|>"
|
| 236 |
-
],
|
| 237 |
"bos_token": null,
|
| 238 |
"clean_up_tokenization_spaces": false,
|
| 239 |
"eos_token": "<|endoftext|>",
|
| 240 |
"errors": "replace",
|
| 241 |
-
"
|
| 242 |
"mask_token": "â½Ĺ",
|
| 243 |
"model_max_length": 131072,
|
| 244 |
"pad_token": "<|endoftext|>",
|
|
|
|
| 1 |
{
|
|
|
|
| 2 |
"add_prefix_space": false,
|
| 3 |
+
"backend": "tokenizers",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
"bos_token": null,
|
| 5 |
"clean_up_tokenization_spaces": false,
|
| 6 |
"eos_token": "<|endoftext|>",
|
| 7 |
"errors": "replace",
|
| 8 |
+
"is_local": true,
|
| 9 |
"mask_token": "â½Ĺ",
|
| 10 |
"model_max_length": 131072,
|
| 11 |
"pad_token": "<|endoftext|>",
|