mirror of
https://github.com/deepseek-ai/DeepSeek-V3.git
synced 2025-02-23 06:08:58 -05:00
b4e06d883e
- Enhanced sentence structure for better clarity and smoother flow. - Adjusted wording and phrasing to improve accuracy and professionalism. - Optimized the organization of information for better readability, especially in the sections related to parameters and technical details. - Refined formatting and sectioning of the documentation for easier navigation and comprehension.
93 lines
3.6 KiB
Markdown
93 lines
3.6 KiB
Markdown
# DeepSeek-V3 Weight File Documentation
|
||
|
||
## New Fields in `config.json`
|
||
|
||
- **model_type**: Specifies the model type, which is now set to `deepseek_v3` in this release.
|
||
- **num_nextn_predict_layers**: Defines the number of Multi-Token Prediction (MTP) Modules. The open-sourced V3 weights contain **1 MTP Module**.
|
||
- **quantization_config**: Details the configuration for FP8 quantization.
|
||
|
||
---
|
||
|
||
## Weight File Structure Overview
|
||
|
||
The DeepSeek-V3 weight file is divided into two primary components: **Main Model Weights** and **MTP Modules**.
|
||
|
||
### 1. Main Model Weights
|
||
|
||
- **Composition**:
|
||
- Includes input/output embedding layers and a full set of 61 Transformer hidden layers.
|
||
- **Parameter Count**:
|
||
- Total parameters: **671B**
|
||
- Activation parameters: **36.7B** (which includes 0.9B for Embedding and 0.9B for the Output Head).
|
||
|
||
#### Structural Details
|
||
|
||
- **Embedding Layer**:
|
||
- `model.embed_tokens.weight`
|
||
- **Transformer Hidden Layers**:
|
||
- From `model.layers.0` to `model.layers.60`, which correspond to `num_hidden_layers` layers.
|
||
- **Output Layer**:
|
||
- `model.norm.weight`
|
||
- `lm_head.weight`
|
||
|
||
### 2. Multi-Token Prediction (MTP) Modules
|
||
|
||
- **Composition**:
|
||
- These modules are determined by the `num_nextn_predict_layers` parameter. In this model, the value is set to 1.
|
||
- **Parameter Count**:
|
||
- Parameters: **11.5B unique parameters** (excluding the shared 0.9B Embedding and 0.9B Output Head).
|
||
- Activation parameters: **2.4B** (including the shared 0.9B Embedding and 0.9B Output Head).
|
||
|
||
#### Structural Details
|
||
|
||
- **embed_tokens**: **Shares parameters** with the Main Model’s Embedding layer.
|
||
- **enorm & hnorm**: RMSNorm parameters used for speculative decoding.
|
||
- **eh_proj**: Parameters used for dimensionality reduction of the normalized outputs.
|
||
- **Additional Transformer Hidden Layer**:
|
||
- `model.layers.61.self_attn & mlp` (these are structured the same as the Main Model hidden layers).
|
||
- **shared_head**: **Shares parameters** with the Output Head of the Main Model.
|
||
|
||
---
|
||
|
||
### Layer Loading Rules
|
||
|
||
- **Main Model Weights**: These are loaded according to the `num_hidden_layers` field in `config.json`.
|
||
- **MTP Modules**: These are loaded using the `num_nextn_predict_layers` field, with MTP layer IDs appended directly after the Main Model’s hidden layers. For example:
|
||
- With `num_hidden_layers = 61` and `num_nextn_predict_layers = 1`, the MTP Module layer ID will be `61`.
|
||
|
||
---
|
||
|
||
## FP8 Weight Documentation
|
||
|
||
DeepSeek-V3 natively supports the FP8 weight format with 128x128 block scaling.
|
||
|
||
### FP8 Configuration
|
||
|
||
The FP8 weight file introduces a `quantization_config` field, which defines the quantization method. Below is an example of the configuration:
|
||
|
||
```json
|
||
"quantization_config": {
|
||
"activation_scheme": "dynamic",
|
||
"fmt": "e4m3",
|
||
"quant_method": "fp8",
|
||
"weight_block_size": [128, 128]
|
||
}
|
||
```
|
||
|
||
- **Quantization Format**:
|
||
- Format type: `fp8` and `e4m3` (aligned with `torch.float8_e4m3fn`).
|
||
- Weight block size: `128x128`.
|
||
- **Activation Quantization Scheme**:
|
||
- Uses dynamic activation quantization (`dynamic`).
|
||
|
||
### Dequantization Method
|
||
|
||
The FP8 weight file includes a `weight_scale_inv` field, which stores the dequantization scale for each weight block.
|
||
|
||
- **Storage Format**: Stored as a `float32 Tensor`, alongside the weight data.
|
||
- **Dequantization Formula**:
|
||
- If a weight block is not aligned to 128, it is zero-padded to 128 before calculating the scale. The padded portion is discarded after quantization.
|
||
- Dequantization is performed using the formula: `(128x128 weight block) * weight_scale_inv`.
|
||
|
||
This dequantization process enables runtime operations to apply online quantization on a per-token, per-128-channel basis.
|