Upload 4-bit quantized version of perplexity-ai/pplx-embed-v1-0.6b with 64.7% memory reduction

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,37 @@
+---
+base_model: perplexity-ai/pplx-embed-v1-0.6b
+language: en
+license: apache-2.0
+tags:
+- quantized
+- 4bit
+- bnb
+- transformers
+model_name: pplx-embed-v1-0.6b-bnb-4bit-nf4
+---
+
+# pplx-embed-v1-0.6b (Quantized)
+
+## Description
+This model is a 4-bit quantized version of the original [`perplexity-ai/pplx-embed-v1-0.6b`](https://huggingface.co/perplexity-ai/pplx-embed-v1-0.6b) model, optimized for reduced memory usage while maintaining performance.
+
+## Quantization Details
+- **Quantization Type**: 4-bit
+- **bnb_4bit_quant_type**: nf4
+- **bnb_4bit_use_double_quant**: True
+- **bnb_4bit_compute_dtype**: bfloat16
+- **bnb_4bit_quant_storage**: uint8
+- **Original Footprint**: 2384.20 MB (FLOAT32)
+- **Quantized Footprint**: 842.79 MB (UINT8)
+- **Memory Reduction**: 64.7%
+
+## Usage
+```python
+from transformers import AutoModel, AutoTokenizer
+
+model_name = "pplx-embed-v1-0.6b-bnb-4bit-nf4"
+model = AutoModel.from_pretrained(
+    "manu02/pplx-embed-v1-0.6b-bnb-4bit-nf4",
+)
+tokenizer = AutoTokenizer.from_pretrained("manu02/pplx-embed-v1-0.6b-bnb-4bit-nf4", use_fast=True)
+```

config.json

@@ -0,0 +1,83 @@
+{
+  "architectures": [
+    "PPLXQwen3Model"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration.PPLXQwen3Config",
+    "AutoModel": "modeling.PPLXQwen3Model"
+  },
+  "bos_token_id": 151643,
+  "dtype": "float32",
+  "eos_token_id": 151643,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 32768,
+  "max_window_layers": 28,
+  "model_type": "bidirectional_pplx_qwen3",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "pad_token_id": null,
+  "quantization_config": {
+    "_load_in_4bit": true,
+    "_load_in_8bit": false,
+    "bnb_4bit_compute_dtype": "bfloat16",
+    "bnb_4bit_quant_storage": "uint8",
+    "bnb_4bit_quant_type": "nf4",
+    "bnb_4bit_use_double_quant": true,
+    "llm_int8_enable_fp32_cpu_offload": false,
+    "llm_int8_has_fp16_weight": false,
+    "llm_int8_skip_modules": null,
+    "llm_int8_threshold": 6.0,
+    "load_in_4bit": true,
+    "load_in_8bit": false,
+    "quant_method": "bitsandbytes"
+  },
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 1000000,
+    "rope_type": "default"
+  },
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.2.0.dev0",
+  "use_bidirectional_attention": true,
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

configuration.py

@@ -0,0 +1,5 @@
+from transformers.models.qwen3.configuration_qwen3 import Qwen3Config
+
+
+class PPLXQwen3Config(Qwen3Config):
+    model_type = "bidirectional_pplx_qwen3"

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccb52a10a1fc0fb4207c742dd145ef35ffa9083cbde8895b3d303152ab27abed
+size 850175239

modeling.py

@@ -0,0 +1,83 @@
+from typing import Callable
+import torch
+from transformers import Qwen3Model
+from transformers.cache_utils import Cache
+from transformers.masking_utils import create_causal_mask
+from transformers.modeling_outputs import BaseModelOutputWithPooling
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs
+from .configuration import PPLXQwen3Config
+
+
+# From modeling_t5gemma.py
+def bidirectional_mask_function(attention_mask: torch.Tensor | None) -> Callable:
+    """
+    This creates bidirectional attention mask.
+    """
+
+    def inner_mask(batch_idx: int, head_idx: int, q_idx: int, kv_idx: int) -> bool:
+        if attention_mask is None:
+            return torch.ones((), dtype=torch.bool)
+        return attention_mask[batch_idx, kv_idx].to(torch.bool)
+
+    return inner_mask
+
+
+class PPLXQwen3Model(Qwen3Model):
+    _supports_flash_attn = True
+    _supports_sdpa = True
+
+    config_class = PPLXQwen3Config
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.post_init()
+
+    def post_init(self):
+        super().post_init()
+        # Override to set all layers to non-causal attention. This'll work with attn_implementation="flash_attention_2" or "sdpa"
+        for layer in self.layers:
+            layer.self_attn.is_causal = False
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        use_cache: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> BaseModelOutputWithPooling:
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+            input_ids = None
+
+        # We construct a dummy tensor imitating initial positions
+        dummy_cache_position = torch.arange(
+            inputs_embeds.shape[1], device=inputs_embeds.device, dtype=torch.long
+        )
+        attention_mask = {
+            "full_attention": create_causal_mask(
+                config=self.config,
+                input_embeds=inputs_embeds,
+                attention_mask=attention_mask,
+                cache_position=dummy_cache_position,
+                past_key_values=None,
+                position_ids=position_ids,
+                or_mask_function=bidirectional_mask_function(attention_mask),
+            )
+        }
+
+        outputs = super().forward(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            **kwargs,
+        )
+        return outputs

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6fb5c5bbba5fa5f8332edfb6d8aa67bd7fb3d75365b1765f108201698eaebf5
+size 11422837

tokenizer_config.json

@@ -0,0 +1,16 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "is_local": false,
+  "mask_token": "â½Ĺ",
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "sep_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "TokenizersBackend",
+  "unk_token": null
+}