Add files using upload-large-folder tool
Browse files- .gitattributes +2 -0
- DeepSeek_V4.pdf +3 -0
- LICENSE +21 -0
- README.md +99 -0
- config.json +429 -0
- encoding/README.md +156 -0
- encoding/encoding_dsv4.py +744 -0
- encoding/test_encoding_dsv4.py +89 -0
- encoding/tests/test_input_1.json +81 -0
- encoding/tests/test_input_2.json +24 -0
- encoding/tests/test_input_3.json +159 -0
- encoding/tests/test_input_4.json +28 -0
- encoding/tests/test_output_1.txt +36 -0
- encoding/tests/test_output_2.txt +1 -0
- encoding/tests/test_output_3.txt +38 -0
- encoding/tests/test_output_4.txt +29 -0
- generation_config.json +12 -0
- jang_config.json +386 -0
- model-00001-of-00082.safetensors +3 -0
- model-00002-of-00082.safetensors +3 -0
- model-00004-of-00082.safetensors +3 -0
- model-00010-of-00082.safetensors +3 -0
- model-00013-of-00082.safetensors +3 -0
- model-00015-of-00082.safetensors +3 -0
- model-00016-of-00082.safetensors +3 -0
- model-00019-of-00082.safetensors +3 -0
- model-00021-of-00082.safetensors +3 -0
- model-00022-of-00082.safetensors +3 -0
- model-00024-of-00082.safetensors +3 -0
- model-00027-of-00082.safetensors +3 -0
- model-00028-of-00082.safetensors +3 -0
- model-00039-of-00082.safetensors +3 -0
- model-00040-of-00082.safetensors +3 -0
- model-00043-of-00082.safetensors +3 -0
- model-00045-of-00082.safetensors +3 -0
- model-00051-of-00082.safetensors +3 -0
- model-00052-of-00082.safetensors +3 -0
- model-00057-of-00082.safetensors +3 -0
- model-00058-of-00082.safetensors +3 -0
- model-00060-of-00082.safetensors +3 -0
- model-00063-of-00082.safetensors +3 -0
- model-00065-of-00082.safetensors +3 -0
- model-00066-of-00082.safetensors +3 -0
- model-00069-of-00082.safetensors +3 -0
- model-00071-of-00082.safetensors +3 -0
- model-00074-of-00082.safetensors +3 -0
- model-00082-of-00082.safetensors +3 -0
- model.safetensors.index.json +0 -0
- tokenizer.json +0 -0
- tokenizer_config.json +38 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
*.pdf filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
*.png filter=lfs diff=lfs merge=lfs -text
|
DeepSeek_V4.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8a03dadae71894de1515da33e296f0df1dbeed3e7f4bf0badd087f9af77f29e9
|
| 3 |
+
size 4480407
|
LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
MIT License
|
| 2 |
+
|
| 3 |
+
Copyright (c) 2023 DeepSeek
|
| 4 |
+
|
| 5 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 6 |
+
of this software and associated documentation files (the "Software"), to deal
|
| 7 |
+
in the Software without restriction, including without limitation the rights
|
| 8 |
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 9 |
+
copies of the Software, and to permit persons to whom the Software is
|
| 10 |
+
furnished to do so, subject to the following conditions:
|
| 11 |
+
|
| 12 |
+
The above copyright notice and this permission notice shall be included in all
|
| 13 |
+
copies or substantial portions of the Software.
|
| 14 |
+
|
| 15 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 16 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 17 |
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 18 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 19 |
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 20 |
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 21 |
+
SOFTWARE.
|
README.md
ADDED
|
@@ -0,0 +1,99 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
- zh
|
| 5 |
+
license: mit
|
| 6 |
+
library_name: mlx
|
| 7 |
+
pipeline_tag: text-generation
|
| 8 |
+
base_model: deepseek-ai/DeepSeek-V4-Flash
|
| 9 |
+
base_model_relation: quantized
|
| 10 |
+
tags:
|
| 11 |
+
- mlx
|
| 12 |
+
- safetensors
|
| 13 |
+
- deepseek_v4
|
| 14 |
+
- deepseek-v4
|
| 15 |
+
- deepseek
|
| 16 |
+
- quantized
|
| 17 |
+
- apple-silicon
|
| 18 |
+
- mixture-of-experts
|
| 19 |
+
- Mixture of Experts
|
| 20 |
+
- mxtq
|
| 21 |
+
- turboquant
|
| 22 |
+
- jang
|
| 23 |
+
- jangtq
|
| 24 |
+
- conversational
|
| 25 |
+
- 2.3-bit
|
| 26 |
+
---
|
| 27 |
+
|
| 28 |
+
# DeepSeek-V4-Flash-JANGTQ-Q2.3-mlx
|
| 29 |
+
|
| 30 |
+
`osmapi/DeepSeek-V4-Flash-JANGTQ-Q2.3-mlx` is an Apple-Silicon MLX TurboQuant/JANGTQ quantization of [`deepseek-ai/DeepSeek-V4-Flash`](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash).
|
| 31 |
+
|
| 32 |
+
No fine-tuning, distillation, or retraining was applied. The official mixed FP4/FP8 source weights were converted locally, the MTP head was dropped because it is not used for normal decode, and router/mHC/control tensors were preserved rather than aggressively quantized.
|
| 33 |
+
|
| 34 |
+
## Model Details
|
| 35 |
+
|
| 36 |
+
| Property | Value |
|
| 37 |
+
|---|---|
|
| 38 |
+
| Base model | `deepseek-ai/DeepSeek-V4-Flash` |
|
| 39 |
+
| Architecture | DeepSeek-V4 Flash MoE, 284B total / 13B active, 1M context |
|
| 40 |
+
| Local profile | `JANGTQ-Q2.3` |
|
| 41 |
+
| Bundle size | 88.03 GB |
|
| 42 |
+
| Layout | Pre-stacked MLX `switch_mlp` layout |
|
| 43 |
+
| MTP head | Dropped |
|
| 44 |
+
| Validation | Safetensors header/index validation, metadata validation |
|
| 45 |
+
|
| 46 |
+
## Required Sidecar
|
| 47 |
+
|
| 48 |
+
This is a JANGTQ/TurboQuant bundle and requires `jangtq_runtime.safetensors` from this repository. The sidecar stores the deterministic codebooks and Hadamard rotation signs used to decode the `.tq_packed` expert weights. If it is missing, re-download the full repository or fetch that file explicitly:
|
| 49 |
+
|
| 50 |
+
```bash
|
| 51 |
+
hf download osmapi/DeepSeek-V4-Flash-JANGTQ-Q2.3-mlx jangtq_runtime.safetensors --local-dir <your-model-dir>
|
| 52 |
+
```
|
| 53 |
+
|
| 54 |
+
## Quantization Recipe
|
| 55 |
+
|
| 56 |
+
| Tensor class | Codec | Bits / handling |
|
| 57 |
+
|---|---:|---|
|
| 58 |
+
| Routed experts | TurboQuant MXTQ | 110 routed layer/projection groups at 2-bit MXTQ and 19 at 4-bit MXTQ |
|
| 59 |
+
| Routed effective bits | MXTQ | 2.2946 bits |
|
| 60 |
+
| Attention, shared experts, compressor, indexer, embed, lm head | MLX affine | 8-bit, group size 32 |
|
| 61 |
+
| Norms, router, mHC, sinks, integer routing tables | passthrough | source precision preserved |
|
| 62 |
+
|
| 63 |
+
The fractional target is implemented as a power-of-two lane mix because the current JANGTQ vectorized packer is stable on 2/4/8-bit lanes for DeepSeek-V4 expert dimensions.
|
| 64 |
+
|
| 65 |
+
## Use
|
| 66 |
+
|
| 67 |
+
Install the JANG loader/runtime and MLX LM:
|
| 68 |
+
|
| 69 |
+
```bash
|
| 70 |
+
pip install mlx-lm jang-tools
|
| 71 |
+
```
|
| 72 |
+
|
| 73 |
+
Example:
|
| 74 |
+
|
| 75 |
+
```python
|
| 76 |
+
from jang_tools.load_jangtq import load_jangtq_model
|
| 77 |
+
from mlx_lm import generate
|
| 78 |
+
|
| 79 |
+
model, tokenizer = load_jangtq_model("osmapi/DeepSeek-V4-Flash-JANGTQ-Q2.3-mlx")
|
| 80 |
+
prompt = "Write a short note about MLX quantization."
|
| 81 |
+
text = generate(model, tokenizer, prompt=prompt, verbose=True)
|
| 82 |
+
print(text)
|
| 83 |
+
```
|
| 84 |
+
|
| 85 |
+
## Files
|
| 86 |
+
|
| 87 |
+
- `model-*.safetensors`: pre-stacked JANGTQ/MLX shards
|
| 88 |
+
- `model.safetensors.index.json`: shard index
|
| 89 |
+
- `jangtq_runtime.safetensors`: required TurboQuant runtime sidecar
|
| 90 |
+
- `config.json`, `jang_config.json`: MLX/JANGTQ metadata
|
| 91 |
+
- `encoding/`: upstream DeepSeek-V4 prompt encoding reference
|
| 92 |
+
|
| 93 |
+
## Notes
|
| 94 |
+
|
| 95 |
+
This card follows the same broad shape as the other osmapi DeepSeek-V4-Flash MLX uploads: a sidecar warning, an explicit recipe table, and minimal reproducible loading instructions. Q2.3 is an aggressive size-first TurboQuant profile, so treat it as experimental until evaluated on your target prompts.
|
| 96 |
+
|
| 97 |
+
## License
|
| 98 |
+
|
| 99 |
+
MIT, following the upstream DeepSeek-V4-Flash release.
|
config.json
ADDED
|
@@ -0,0 +1,429 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"DeepseekV4ForCausalLM"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"bos_token_id": 0,
|
| 8 |
+
"eos_token_id": [
|
| 9 |
+
1,
|
| 10 |
+
128803
|
| 11 |
+
],
|
| 12 |
+
"expert_dtype": "fp4",
|
| 13 |
+
"hc_eps": 1e-06,
|
| 14 |
+
"hc_mult": 4,
|
| 15 |
+
"hc_sinkhorn_iters": 20,
|
| 16 |
+
"head_dim": 512,
|
| 17 |
+
"hidden_act": "silu",
|
| 18 |
+
"hidden_size": 4096,
|
| 19 |
+
"index_head_dim": 128,
|
| 20 |
+
"index_n_heads": 64,
|
| 21 |
+
"index_topk": 512,
|
| 22 |
+
"initializer_range": 0.02,
|
| 23 |
+
"max_position_embeddings": 1048576,
|
| 24 |
+
"model_type": "deepseek_v4",
|
| 25 |
+
"moe_intermediate_size": 2048,
|
| 26 |
+
"n_routed_experts": 256,
|
| 27 |
+
"n_shared_experts": 1,
|
| 28 |
+
"norm_topk_prob": true,
|
| 29 |
+
"num_attention_heads": 64,
|
| 30 |
+
"num_experts_per_tok": 6,
|
| 31 |
+
"num_hidden_layers": 43,
|
| 32 |
+
"num_hash_layers": 3,
|
| 33 |
+
"num_key_value_heads": 1,
|
| 34 |
+
"num_nextn_predict_layers": 0,
|
| 35 |
+
"o_groups": 8,
|
| 36 |
+
"o_lora_rank": 1024,
|
| 37 |
+
"q_lora_rank": 1024,
|
| 38 |
+
"qk_rope_head_dim": 64,
|
| 39 |
+
"rms_norm_eps": 1e-06,
|
| 40 |
+
"rope_theta": 10000,
|
| 41 |
+
"routed_scaling_factor": 1.5,
|
| 42 |
+
"scoring_func": "sqrtsoftplus",
|
| 43 |
+
"sliding_window": 128,
|
| 44 |
+
"swiglu_limit": 10.0,
|
| 45 |
+
"tie_word_embeddings": false,
|
| 46 |
+
"topk_method": "noaux_tc",
|
| 47 |
+
"torch_dtype": "bfloat16",
|
| 48 |
+
"transformers_version": "4.57.1",
|
| 49 |
+
"use_cache": true,
|
| 50 |
+
"vocab_size": 129280,
|
| 51 |
+
"compress_rope_theta": 160000,
|
| 52 |
+
"compress_ratios": [
|
| 53 |
+
0,
|
| 54 |
+
0,
|
| 55 |
+
4,
|
| 56 |
+
128,
|
| 57 |
+
4,
|
| 58 |
+
128,
|
| 59 |
+
4,
|
| 60 |
+
128,
|
| 61 |
+
4,
|
| 62 |
+
128,
|
| 63 |
+
4,
|
| 64 |
+
128,
|
| 65 |
+
4,
|
| 66 |
+
128,
|
| 67 |
+
4,
|
| 68 |
+
128,
|
| 69 |
+
4,
|
| 70 |
+
128,
|
| 71 |
+
4,
|
| 72 |
+
128,
|
| 73 |
+
4,
|
| 74 |
+
128,
|
| 75 |
+
4,
|
| 76 |
+
128,
|
| 77 |
+
4,
|
| 78 |
+
128,
|
| 79 |
+
4,
|
| 80 |
+
128,
|
| 81 |
+
4,
|
| 82 |
+
128,
|
| 83 |
+
4,
|
| 84 |
+
128,
|
| 85 |
+
4,
|
| 86 |
+
128,
|
| 87 |
+
4,
|
| 88 |
+
128,
|
| 89 |
+
4,
|
| 90 |
+
128,
|
| 91 |
+
4,
|
| 92 |
+
128,
|
| 93 |
+
4,
|
| 94 |
+
128,
|
| 95 |
+
4,
|
| 96 |
+
0
|
| 97 |
+
],
|
| 98 |
+
"rope_parameters": {
|
| 99 |
+
"beta_fast": 32.0,
|
| 100 |
+
"beta_slow": 1.0,
|
| 101 |
+
"factor": 16.0,
|
| 102 |
+
"original_max_position_embeddings": 65536,
|
| 103 |
+
"rope_type": "yarn",
|
| 104 |
+
"rope_theta": 10000.0
|
| 105 |
+
},
|
| 106 |
+
"quantization": {
|
| 107 |
+
"bits": 8,
|
| 108 |
+
"group_size": 32,
|
| 109 |
+
"mode": "affine",
|
| 110 |
+
"routed_expert_bits": 2,
|
| 111 |
+
"routed_expert_bit_plan": {
|
| 112 |
+
"codec": "mxtq",
|
| 113 |
+
"target": "q2.3",
|
| 114 |
+
"effective_bits": 2.2946,
|
| 115 |
+
"bits_used": [
|
| 116 |
+
2,
|
| 117 |
+
4
|
| 118 |
+
],
|
| 119 |
+
"group_count": 129,
|
| 120 |
+
"bit_counts": {
|
| 121 |
+
"2": 110,
|
| 122 |
+
"4": 19
|
| 123 |
+
},
|
| 124 |
+
"assignment": {
|
| 125 |
+
"layer_0.down_proj": 4,
|
| 126 |
+
"layer_0.gate_proj": 4,
|
| 127 |
+
"layer_0.up_proj": 4,
|
| 128 |
+
"layer_1.down_proj": 4,
|
| 129 |
+
"layer_1.gate_proj": 4,
|
| 130 |
+
"layer_1.up_proj": 4,
|
| 131 |
+
"layer_2.down_proj": 4,
|
| 132 |
+
"layer_2.gate_proj": 4,
|
| 133 |
+
"layer_2.up_proj": 4,
|
| 134 |
+
"layer_3.down_proj": 4,
|
| 135 |
+
"layer_3.gate_proj": 2,
|
| 136 |
+
"layer_3.up_proj": 2,
|
| 137 |
+
"layer_4.down_proj": 4,
|
| 138 |
+
"layer_4.gate_proj": 2,
|
| 139 |
+
"layer_4.up_proj": 2,
|
| 140 |
+
"layer_5.down_proj": 4,
|
| 141 |
+
"layer_5.gate_proj": 2,
|
| 142 |
+
"layer_5.up_proj": 2,
|
| 143 |
+
"layer_6.down_proj": 4,
|
| 144 |
+
"layer_6.gate_proj": 2,
|
| 145 |
+
"layer_6.up_proj": 2,
|
| 146 |
+
"layer_7.down_proj": 4,
|
| 147 |
+
"layer_7.gate_proj": 2,
|
| 148 |
+
"layer_7.up_proj": 2,
|
| 149 |
+
"layer_8.down_proj": 4,
|
| 150 |
+
"layer_8.gate_proj": 2,
|
| 151 |
+
"layer_8.up_proj": 2,
|
| 152 |
+
"layer_9.down_proj": 4,
|
| 153 |
+
"layer_9.gate_proj": 2,
|
| 154 |
+
"layer_9.up_proj": 2,
|
| 155 |
+
"layer_10.down_proj": 4,
|
| 156 |
+
"layer_10.gate_proj": 2,
|
| 157 |
+
"layer_10.up_proj": 2,
|
| 158 |
+
"layer_11.down_proj": 4,
|
| 159 |
+
"layer_11.gate_proj": 2,
|
| 160 |
+
"layer_11.up_proj": 2,
|
| 161 |
+
"layer_12.down_proj": 4,
|
| 162 |
+
"layer_12.gate_proj": 2,
|
| 163 |
+
"layer_12.up_proj": 2,
|
| 164 |
+
"layer_13.down_proj": 2,
|
| 165 |
+
"layer_13.gate_proj": 2,
|
| 166 |
+
"layer_13.up_proj": 2,
|
| 167 |
+
"layer_14.down_proj": 2,
|
| 168 |
+
"layer_14.gate_proj": 2,
|
| 169 |
+
"layer_14.up_proj": 2,
|
| 170 |
+
"layer_15.down_proj": 2,
|
| 171 |
+
"layer_15.gate_proj": 2,
|
| 172 |
+
"layer_15.up_proj": 2,
|
| 173 |
+
"layer_16.down_proj": 2,
|
| 174 |
+
"layer_16.gate_proj": 2,
|
| 175 |
+
"layer_16.up_proj": 2,
|
| 176 |
+
"layer_17.down_proj": 2,
|
| 177 |
+
"layer_17.gate_proj": 2,
|
| 178 |
+
"layer_17.up_proj": 2,
|
| 179 |
+
"layer_18.down_proj": 2,
|
| 180 |
+
"layer_18.gate_proj": 2,
|
| 181 |
+
"layer_18.up_proj": 2,
|
| 182 |
+
"layer_19.down_proj": 2,
|
| 183 |
+
"layer_19.gate_proj": 2,
|
| 184 |
+
"layer_19.up_proj": 2,
|
| 185 |
+
"layer_20.down_proj": 2,
|
| 186 |
+
"layer_20.gate_proj": 2,
|
| 187 |
+
"layer_20.up_proj": 2,
|
| 188 |
+
"layer_21.down_proj": 2,
|
| 189 |
+
"layer_21.gate_proj": 2,
|
| 190 |
+
"layer_21.up_proj": 2,
|
| 191 |
+
"layer_22.down_proj": 2,
|
| 192 |
+
"layer_22.gate_proj": 2,
|
| 193 |
+
"layer_22.up_proj": 2,
|
| 194 |
+
"layer_23.down_proj": 2,
|
| 195 |
+
"layer_23.gate_proj": 2,
|
| 196 |
+
"layer_23.up_proj": 2,
|
| 197 |
+
"layer_24.down_proj": 2,
|
| 198 |
+
"layer_24.gate_proj": 2,
|
| 199 |
+
"layer_24.up_proj": 2,
|
| 200 |
+
"layer_25.down_proj": 2,
|
| 201 |
+
"layer_25.gate_proj": 2,
|
| 202 |
+
"layer_25.up_proj": 2,
|
| 203 |
+
"layer_26.down_proj": 2,
|
| 204 |
+
"layer_26.gate_proj": 2,
|
| 205 |
+
"layer_26.up_proj": 2,
|
| 206 |
+
"layer_27.down_proj": 2,
|
| 207 |
+
"layer_27.gate_proj": 2,
|
| 208 |
+
"layer_27.up_proj": 2,
|
| 209 |
+
"layer_28.down_proj": 2,
|
| 210 |
+
"layer_28.gate_proj": 2,
|
| 211 |
+
"layer_28.up_proj": 2,
|
| 212 |
+
"layer_29.down_proj": 2,
|
| 213 |
+
"layer_29.gate_proj": 2,
|
| 214 |
+
"layer_29.up_proj": 2,
|
| 215 |
+
"layer_30.down_proj": 2,
|
| 216 |
+
"layer_30.gate_proj": 2,
|
| 217 |
+
"layer_30.up_proj": 2,
|
| 218 |
+
"layer_31.down_proj": 2,
|
| 219 |
+
"layer_31.gate_proj": 2,
|
| 220 |
+
"layer_31.up_proj": 2,
|
| 221 |
+
"layer_32.down_proj": 2,
|
| 222 |
+
"layer_32.gate_proj": 2,
|
| 223 |
+
"layer_32.up_proj": 2,
|
| 224 |
+
"layer_33.down_proj": 2,
|
| 225 |
+
"layer_33.gate_proj": 2,
|
| 226 |
+
"layer_33.up_proj": 2,
|
| 227 |
+
"layer_34.down_proj": 2,
|
| 228 |
+
"layer_34.gate_proj": 2,
|
| 229 |
+
"layer_34.up_proj": 2,
|
| 230 |
+
"layer_35.down_proj": 2,
|
| 231 |
+
"layer_35.gate_proj": 2,
|
| 232 |
+
"layer_35.up_proj": 2,
|
| 233 |
+
"layer_36.down_proj": 2,
|
| 234 |
+
"layer_36.gate_proj": 2,
|
| 235 |
+
"layer_36.up_proj": 2,
|
| 236 |
+
"layer_37.down_proj": 2,
|
| 237 |
+
"layer_37.gate_proj": 2,
|
| 238 |
+
"layer_37.up_proj": 2,
|
| 239 |
+
"layer_38.down_proj": 2,
|
| 240 |
+
"layer_38.gate_proj": 2,
|
| 241 |
+
"layer_38.up_proj": 2,
|
| 242 |
+
"layer_39.down_proj": 2,
|
| 243 |
+
"layer_39.gate_proj": 2,
|
| 244 |
+
"layer_39.up_proj": 2,
|
| 245 |
+
"layer_40.down_proj": 2,
|
| 246 |
+
"layer_40.gate_proj": 2,
|
| 247 |
+
"layer_40.up_proj": 2,
|
| 248 |
+
"layer_41.down_proj": 2,
|
| 249 |
+
"layer_41.gate_proj": 2,
|
| 250 |
+
"layer_41.up_proj": 2,
|
| 251 |
+
"layer_42.down_proj": 2,
|
| 252 |
+
"layer_42.gate_proj": 2,
|
| 253 |
+
"layer_42.up_proj": 2
|
| 254 |
+
},
|
| 255 |
+
"assignment_unit": "layer.projection, shared across all 256 routed experts"
|
| 256 |
+
},
|
| 257 |
+
"mxtq_bits": {
|
| 258 |
+
"routed_expert": 2,
|
| 259 |
+
"attention": 8,
|
| 260 |
+
"shared_expert": 8,
|
| 261 |
+
"compressor": 8,
|
| 262 |
+
"indexer": 8,
|
| 263 |
+
"embed_tokens": 8,
|
| 264 |
+
"lm_head": 8,
|
| 265 |
+
"norms_router_hc": 16
|
| 266 |
+
}
|
| 267 |
+
},
|
| 268 |
+
"weight_format": "mxtq",
|
| 269 |
+
"routed_expert_bits": 2,
|
| 270 |
+
"routed_expert_bit_plan": {
|
| 271 |
+
"codec": "mxtq",
|
| 272 |
+
"target": "q2.3",
|
| 273 |
+
"effective_bits": 2.2946,
|
| 274 |
+
"bits_used": [
|
| 275 |
+
2,
|
| 276 |
+
4
|
| 277 |
+
],
|
| 278 |
+
"group_count": 129,
|
| 279 |
+
"bit_counts": {
|
| 280 |
+
"2": 110,
|
| 281 |
+
"4": 19
|
| 282 |
+
},
|
| 283 |
+
"assignment": {
|
| 284 |
+
"layer_0.down_proj": 4,
|
| 285 |
+
"layer_0.gate_proj": 4,
|
| 286 |
+
"layer_0.up_proj": 4,
|
| 287 |
+
"layer_1.down_proj": 4,
|
| 288 |
+
"layer_1.gate_proj": 4,
|
| 289 |
+
"layer_1.up_proj": 4,
|
| 290 |
+
"layer_2.down_proj": 4,
|
| 291 |
+
"layer_2.gate_proj": 4,
|
| 292 |
+
"layer_2.up_proj": 4,
|
| 293 |
+
"layer_3.down_proj": 4,
|
| 294 |
+
"layer_3.gate_proj": 2,
|
| 295 |
+
"layer_3.up_proj": 2,
|
| 296 |
+
"layer_4.down_proj": 4,
|
| 297 |
+
"layer_4.gate_proj": 2,
|
| 298 |
+
"layer_4.up_proj": 2,
|
| 299 |
+
"layer_5.down_proj": 4,
|
| 300 |
+
"layer_5.gate_proj": 2,
|
| 301 |
+
"layer_5.up_proj": 2,
|
| 302 |
+
"layer_6.down_proj": 4,
|
| 303 |
+
"layer_6.gate_proj": 2,
|
| 304 |
+
"layer_6.up_proj": 2,
|
| 305 |
+
"layer_7.down_proj": 4,
|
| 306 |
+
"layer_7.gate_proj": 2,
|
| 307 |
+
"layer_7.up_proj": 2,
|
| 308 |
+
"layer_8.down_proj": 4,
|
| 309 |
+
"layer_8.gate_proj": 2,
|
| 310 |
+
"layer_8.up_proj": 2,
|
| 311 |
+
"layer_9.down_proj": 4,
|
| 312 |
+
"layer_9.gate_proj": 2,
|
| 313 |
+
"layer_9.up_proj": 2,
|
| 314 |
+
"layer_10.down_proj": 4,
|
| 315 |
+
"layer_10.gate_proj": 2,
|
| 316 |
+
"layer_10.up_proj": 2,
|
| 317 |
+
"layer_11.down_proj": 4,
|
| 318 |
+
"layer_11.gate_proj": 2,
|
| 319 |
+
"layer_11.up_proj": 2,
|
| 320 |
+
"layer_12.down_proj": 4,
|
| 321 |
+
"layer_12.gate_proj": 2,
|
| 322 |
+
"layer_12.up_proj": 2,
|
| 323 |
+
"layer_13.down_proj": 2,
|
| 324 |
+
"layer_13.gate_proj": 2,
|
| 325 |
+
"layer_13.up_proj": 2,
|
| 326 |
+
"layer_14.down_proj": 2,
|
| 327 |
+
"layer_14.gate_proj": 2,
|
| 328 |
+
"layer_14.up_proj": 2,
|
| 329 |
+
"layer_15.down_proj": 2,
|
| 330 |
+
"layer_15.gate_proj": 2,
|
| 331 |
+
"layer_15.up_proj": 2,
|
| 332 |
+
"layer_16.down_proj": 2,
|
| 333 |
+
"layer_16.gate_proj": 2,
|
| 334 |
+
"layer_16.up_proj": 2,
|
| 335 |
+
"layer_17.down_proj": 2,
|
| 336 |
+
"layer_17.gate_proj": 2,
|
| 337 |
+
"layer_17.up_proj": 2,
|
| 338 |
+
"layer_18.down_proj": 2,
|
| 339 |
+
"layer_18.gate_proj": 2,
|
| 340 |
+
"layer_18.up_proj": 2,
|
| 341 |
+
"layer_19.down_proj": 2,
|
| 342 |
+
"layer_19.gate_proj": 2,
|
| 343 |
+
"layer_19.up_proj": 2,
|
| 344 |
+
"layer_20.down_proj": 2,
|
| 345 |
+
"layer_20.gate_proj": 2,
|
| 346 |
+
"layer_20.up_proj": 2,
|
| 347 |
+
"layer_21.down_proj": 2,
|
| 348 |
+
"layer_21.gate_proj": 2,
|
| 349 |
+
"layer_21.up_proj": 2,
|
| 350 |
+
"layer_22.down_proj": 2,
|
| 351 |
+
"layer_22.gate_proj": 2,
|
| 352 |
+
"layer_22.up_proj": 2,
|
| 353 |
+
"layer_23.down_proj": 2,
|
| 354 |
+
"layer_23.gate_proj": 2,
|
| 355 |
+
"layer_23.up_proj": 2,
|
| 356 |
+
"layer_24.down_proj": 2,
|
| 357 |
+
"layer_24.gate_proj": 2,
|
| 358 |
+
"layer_24.up_proj": 2,
|
| 359 |
+
"layer_25.down_proj": 2,
|
| 360 |
+
"layer_25.gate_proj": 2,
|
| 361 |
+
"layer_25.up_proj": 2,
|
| 362 |
+
"layer_26.down_proj": 2,
|
| 363 |
+
"layer_26.gate_proj": 2,
|
| 364 |
+
"layer_26.up_proj": 2,
|
| 365 |
+
"layer_27.down_proj": 2,
|
| 366 |
+
"layer_27.gate_proj": 2,
|
| 367 |
+
"layer_27.up_proj": 2,
|
| 368 |
+
"layer_28.down_proj": 2,
|
| 369 |
+
"layer_28.gate_proj": 2,
|
| 370 |
+
"layer_28.up_proj": 2,
|
| 371 |
+
"layer_29.down_proj": 2,
|
| 372 |
+
"layer_29.gate_proj": 2,
|
| 373 |
+
"layer_29.up_proj": 2,
|
| 374 |
+
"layer_30.down_proj": 2,
|
| 375 |
+
"layer_30.gate_proj": 2,
|
| 376 |
+
"layer_30.up_proj": 2,
|
| 377 |
+
"layer_31.down_proj": 2,
|
| 378 |
+
"layer_31.gate_proj": 2,
|
| 379 |
+
"layer_31.up_proj": 2,
|
| 380 |
+
"layer_32.down_proj": 2,
|
| 381 |
+
"layer_32.gate_proj": 2,
|
| 382 |
+
"layer_32.up_proj": 2,
|
| 383 |
+
"layer_33.down_proj": 2,
|
| 384 |
+
"layer_33.gate_proj": 2,
|
| 385 |
+
"layer_33.up_proj": 2,
|
| 386 |
+
"layer_34.down_proj": 2,
|
| 387 |
+
"layer_34.gate_proj": 2,
|
| 388 |
+
"layer_34.up_proj": 2,
|
| 389 |
+
"layer_35.down_proj": 2,
|
| 390 |
+
"layer_35.gate_proj": 2,
|
| 391 |
+
"layer_35.up_proj": 2,
|
| 392 |
+
"layer_36.down_proj": 2,
|
| 393 |
+
"layer_36.gate_proj": 2,
|
| 394 |
+
"layer_36.up_proj": 2,
|
| 395 |
+
"layer_37.down_proj": 2,
|
| 396 |
+
"layer_37.gate_proj": 2,
|
| 397 |
+
"layer_37.up_proj": 2,
|
| 398 |
+
"layer_38.down_proj": 2,
|
| 399 |
+
"layer_38.gate_proj": 2,
|
| 400 |
+
"layer_38.up_proj": 2,
|
| 401 |
+
"layer_39.down_proj": 2,
|
| 402 |
+
"layer_39.gate_proj": 2,
|
| 403 |
+
"layer_39.up_proj": 2,
|
| 404 |
+
"layer_40.down_proj": 2,
|
| 405 |
+
"layer_40.gate_proj": 2,
|
| 406 |
+
"layer_40.up_proj": 2,
|
| 407 |
+
"layer_41.down_proj": 2,
|
| 408 |
+
"layer_41.gate_proj": 2,
|
| 409 |
+
"layer_41.up_proj": 2,
|
| 410 |
+
"layer_42.down_proj": 2,
|
| 411 |
+
"layer_42.gate_proj": 2,
|
| 412 |
+
"layer_42.up_proj": 2
|
| 413 |
+
},
|
| 414 |
+
"assignment_unit": "layer.projection, shared across all 256 routed experts"
|
| 415 |
+
},
|
| 416 |
+
"mxtq_bits": {
|
| 417 |
+
"routed_expert": 2,
|
| 418 |
+
"attention": 8,
|
| 419 |
+
"shared_expert": 8,
|
| 420 |
+
"compressor": 8,
|
| 421 |
+
"indexer": 8,
|
| 422 |
+
"embed_tokens": 8,
|
| 423 |
+
"lm_head": 8,
|
| 424 |
+
"norms_router_hc": 16
|
| 425 |
+
},
|
| 426 |
+
"mxtq_seed": 42,
|
| 427 |
+
"group_size": 32,
|
| 428 |
+
"_name_or_path": "DSV4F-MLX-TQ-Q2.3"
|
| 429 |
+
}
|
encoding/README.md
ADDED
|
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# DeepSeek-V4 Encoding
|
| 2 |
+
|
| 3 |
+
This document describes the prompt encoding format used by DeepSeek-V4 series models. The encoding handles multi-turn conversations, tool calling, extended thinking (reasoning), and quick instruction tasks.
|
| 4 |
+
|
| 5 |
+
A self-contained reference implementation is provided in `encoding_dsv4.py`.
|
| 6 |
+
|
| 7 |
+
## Quick Start
|
| 8 |
+
|
| 9 |
+
```python
|
| 10 |
+
from encoding_dsv4 import encode_messages, parse_message_from_completion_text
|
| 11 |
+
|
| 12 |
+
# Encode a conversation
|
| 13 |
+
messages = [
|
| 14 |
+
{"role": "system", "content": "You are a helpful assistant."},
|
| 15 |
+
{"role": "user", "content": "What is 2+2?"},
|
| 16 |
+
]
|
| 17 |
+
prompt = encode_messages(messages, thinking_mode="thinking")
|
| 18 |
+
# => "<｜begin▁of▁sentence｜>You are a helpful assistant.<｜User｜>What is 2+2?<｜Assistant｜><think>"
|
| 19 |
+
|
| 20 |
+
# Parse model output back to structured message
|
| 21 |
+
completion = "Simple arithmetic.</think>2 + 2 = 4.<｜end▁of▁sentence｜>"
|
| 22 |
+
parsed = parse_message_from_completion_text(completion, thinking_mode="thinking")
|
| 23 |
+
# => {"role": "assistant", "reasoning_content": "Simple arithmetic.", "content": "2 + 2 = 4.", "tool_calls": []}
|
| 24 |
+
```
|
| 25 |
+
|
| 26 |
+
> **Note:** The `parse_message_from_completion_text` function is designed to handle well-formatted model output only. It does not attempt to correct or recover from malformed output that the model might occasionally generate. For production use, additional error handling is recommended.
|
| 27 |
+
|
| 28 |
+
## Message Format
|
| 29 |
+
|
| 30 |
+
### Special Tokens
|
| 31 |
+
|
| 32 |
+
| Token | Purpose |
|
| 33 |
+
|-------|---------|
|
| 34 |
+
| `<｜begin▁of▁sentence｜>` | Beginning of sequence (BOS) |
|
| 35 |
+
| `<｜end▁of▁sentence｜>` | End of assistant turn (EOS) |
|
| 36 |
+
| `<｜User｜>` | User turn prefix |
|
| 37 |
+
| `<｜Assistant｜>` | Assistant turn prefix |
|
| 38 |
+
| `<｜latest_reminder｜>` | Latest reminder (date, locale, etc.) |
|
| 39 |
+
| `<think>` / `</think>` | Reasoning block delimiters |
|
| 40 |
+
| `｜DSML｜` | DSML markup token |
|
| 41 |
+
|
| 42 |
+
### Roles
|
| 43 |
+
|
| 44 |
+
The encoding supports the following message roles: `system`, `user`, `assistant`, `tool`, `latest_reminder`, and `developer`.
|
| 45 |
+
|
| 46 |
+
> **Note on the `developer` role:** The `developer` role is used exclusively in the internal search agent pipeline. It is not needed for general-purpose chat or tool-calling tasks, and the official API does not accept messages with this role.
|
| 47 |
+
|
| 48 |
+
### Basic Chat
|
| 49 |
+
|
| 50 |
+
A simple multi-turn conversation is encoded as:
|
| 51 |
+
|
| 52 |
+
```
|
| 53 |
+
<｜begin▁of▁sentence｜>{system_prompt}
|
| 54 |
+
<｜User｜>{user_message}<｜Assistant｜></think>{response}<｜end▁of▁sentence｜>
|
| 55 |
+
<｜User｜>{user_message_2}<｜Assistant｜></think>{response_2}<｜end▁of▁sentence｜>
|
| 56 |
+
```
|
| 57 |
+
|
| 58 |
+
- The BOS token is prepended at the very beginning of the conversation.
|
| 59 |
+
- In **chat mode** (`thinking_mode="chat"`), `</think>` is placed right after `<｜Assistant｜>` to immediately close the thinking block, so the model generates content directly.
|
| 60 |
+
|
| 61 |
+
### Interleaved Thinking Mode
|
| 62 |
+
|
| 63 |
+
In **thinking mode** (`thinking_mode="thinking"`), the model produces explicit reasoning inside `<think>...</think>` blocks before responding.
|
| 64 |
+
|
| 65 |
+
```
|
| 66 |
+
<｜begin▁of▁sentence｜>{system_prompt}
|
| 67 |
+
<｜User｜>{message}<｜Assistant｜><think>{reasoning}</think>{response}<｜end▁of▁sentence｜>
|
| 68 |
+
```
|
| 69 |
+
|
| 70 |
+
The `drop_thinking` parameter (default `True`) controls whether reasoning from earlier turns is preserved:
|
| 71 |
+
|
| 72 |
+
- **Without tools**: `drop_thinking` takes effect. Reasoning content from assistant turns **before** the last user message is stripped. Only the final assistant turn retains its `<think>...</think>` block.
|
| 73 |
+
- **With tools** (on system or developer message): `drop_thinking` is automatically disabled. All turns retain their reasoning, because tool-calling conversations require full context for the model to track multi-step reasoning across tool calls.
|
| 74 |
+
|
| 75 |
+
### Tool Calling (DSML Format)
|
| 76 |
+
|
| 77 |
+
Tools are defined on the `system` or `developer` message via the `tools` field (OpenAI-compatible format). When tools are present, the following schema block is injected into the system/user prompt:
|
| 78 |
+
|
| 79 |
+
```
|
| 80 |
+
## Tools
|
| 81 |
+
|
| 82 |
+
You have access to a set of tools to help answer the user's question. You can invoke tools by writing a "<｜DSML｜tool_calls>" block like the following:
|
| 83 |
+
|
| 84 |
+
<｜DSML｜tool_calls>
|
| 85 |
+
<｜DSML｜invoke name="$TOOL_NAME">
|
| 86 |
+
<｜DSML｜parameter name="$PARAMETER_NAME" string="true|false">$PARAMETER_VALUE</｜DSML｜parameter>
|
| 87 |
+
...
|
| 88 |
+
</｜DSML｜invoke>
|
| 89 |
+
<｜DSML｜invoke name="$TOOL_NAME2">
|
| 90 |
+
...
|
| 91 |
+
</｜DSML｜invoke>
|
| 92 |
+
</｜DSML｜tool_calls>
|
| 93 |
+
|
| 94 |
+
String parameters should be specified as is and set `string="true"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string="false"`.
|
| 95 |
+
|
| 96 |
+
If thinking_mode is enabled (triggered by <think>), you MUST output your complete reasoning inside <think>...</think> BEFORE any tool calls or final response.
|
| 97 |
+
|
| 98 |
+
Otherwise, output directly after </think> with tool calls or final response.
|
| 99 |
+
|
| 100 |
+
### Available Tool Schemas
|
| 101 |
+
|
| 102 |
+
{tool_definitions_json}
|
| 103 |
+
|
| 104 |
+
You MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.
|
| 105 |
+
```
|
| 106 |
+
|
| 107 |
+
An actual tool call in the assistant turn looks like:
|
| 108 |
+
|
| 109 |
+
```xml
|
| 110 |
+
<｜DSML｜tool_calls>
|
| 111 |
+
<｜DSML｜invoke name="function_name">
|
| 112 |
+
<｜DSML｜parameter name="param" string="true">string_value</｜DSML｜parameter>
|
| 113 |
+
<｜DSML｜parameter name="count" string="false">5</｜DSML｜parameter>
|
| 114 |
+
</｜DSML｜invoke>
|
| 115 |
+
</｜DSML｜tool_calls><｜end▁of▁sentence｜>
|
| 116 |
+
```
|
| 117 |
+
|
| 118 |
+
- `string="true"`: the parameter value is a raw string.
|
| 119 |
+
- `string="false"`: the parameter value is JSON (number, boolean, array, object).
|
| 120 |
+
|
| 121 |
+
Tool execution results are wrapped in `<tool_result>` tags within user messages:
|
| 122 |
+
|
| 123 |
+
```
|
| 124 |
+
<｜User｜><tool_result>{result_json}</tool_result><｜Assistant｜><think>...
|
| 125 |
+
```
|
| 126 |
+
|
| 127 |
+
When multiple tool results are present, they are sorted by the order of the corresponding `tool_calls` in the preceding assistant message.
|
| 128 |
+
|
| 129 |
+
### Reasoning Effort
|
| 130 |
+
|
| 131 |
+
When `reasoning_effort="max"` is set, a special prefix is prepended at the very beginning of the prompt (before the system message) to instruct the model to maximize its reasoning depth:
|
| 132 |
+
|
| 133 |
+
```
|
| 134 |
+
Reasoning Effort: Absolute maximum with no shortcuts permitted.
|
| 135 |
+
You MUST be very thorough in your thinking and comprehensively decompose the problem to resolve the root cause, rigorously stress-testing your logic against all potential paths, edge cases, and adversarial scenarios.
|
| 136 |
+
Explicitly write out your entire deliberation process, documenting every intermediate step, considered alternative, and rejected hypothesis to ensure absolutely no assumption is left unchecked.
|
| 137 |
+
```
|
| 138 |
+
|
| 139 |
+
### Quick Instruction Special Tokens
|
| 140 |
+
|
| 141 |
+
Quick instruction tokens are used for auxiliary classification and generation tasks. They are appended to messages via the `"task"` field to trigger specialized model behavior for a single-token or short-form output.
|
| 142 |
+
|
| 143 |
+
| Special Token | Description | Format |
|
| 144 |
+
|:---|:---|:---|
|
| 145 |
+
| `<｜action｜>` | Determines whether the user prompt requires a web search or can be answered directly. | `...<｜User｜>{prompt}<｜Assistant｜><think><｜action｜>` |
|
| 146 |
+
| `<｜title｜>` | Generates a concise conversation title after the first assistant response. | `...<｜Assistant｜>{response}<｜end▁of▁sentence｜><｜title｜>` |
|
| 147 |
+
| `<｜query｜>` | Generates search queries for the user prompt. | `...<｜User｜>{prompt}<｜query｜>` |
|
| 148 |
+
| `<｜authority｜>` | Classifies the user prompt's demand for source authoritativeness. | `...<｜User｜>{prompt}<｜authority｜>` |
|
| 149 |
+
| `<｜domain｜>` | Identifies the domain of the user prompt. | `...<｜User｜>{prompt}<｜domain｜>` |
|
| 150 |
+
| `<｜extracted_url｜>` `<｜read_url｜>` | Determines whether each URL in the user prompt should be fetched and read. | `...<｜User｜>{prompt}<｜extracted_url｜>{url}<｜read_url｜>` |
|
| 151 |
+
|
| 152 |
+
Usage in message format:
|
| 153 |
+
|
| 154 |
+
- **`action`** on a user message: the `<｜action｜>` token is placed after the assistant prefix and thinking token, triggering a routing decision (e.g., "Search" or "Answer").
|
| 155 |
+
- **Other tasks** (`query`, `authority`, `domain`, `read_url`) on a user message: the task token is appended directly after the user content.
|
| 156 |
+
- **`title`** on an assistant message: the `<｜title｜>` token is appended after the assistant's EOS. The next assistant message provides the generated title.
|
encoding/encoding_dsv4.py
ADDED
|
@@ -0,0 +1,744 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
DeepSeek-V4 Encoding
|
| 3 |
+
|
| 4 |
+
A self-contained implementation for encoding/decoding DeepSeek-V4 chat messages
|
| 5 |
+
with tool calling, thinking mode, and quick instruction task support.
|
| 6 |
+
"""
|
| 7 |
+
|
| 8 |
+
from typing import Any, Dict, List, Union, Optional, Tuple
|
| 9 |
+
import copy
|
| 10 |
+
import json
|
| 11 |
+
import re
|
| 12 |
+
|
| 13 |
+
# ============================================================
|
| 14 |
+
# Special Tokens
|
| 15 |
+
# ============================================================
|
| 16 |
+
|
| 17 |
+
bos_token: str = "<｜begin▁of▁sentence｜>"
|
| 18 |
+
eos_token: str = "<｜end▁of▁sentence｜>"
|
| 19 |
+
thinking_start_token: str = "<think>"
|
| 20 |
+
thinking_end_token: str = "</think>"
|
| 21 |
+
dsml_token: str = "｜DSML｜"
|
| 22 |
+
|
| 23 |
+
USER_SP_TOKEN = "<｜User｜>"
|
| 24 |
+
ASSISTANT_SP_TOKEN = "<｜Assistant｜>"
|
| 25 |
+
LATEST_REMINDER_SP_TOKEN = "<｜latest_reminder｜>"
|
| 26 |
+
|
| 27 |
+
# Task special tokens for internal classification tasks
|
| 28 |
+
DS_TASK_SP_TOKENS = {
|
| 29 |
+
"action": "<｜action｜>",
|
| 30 |
+
"query": "<｜query｜>",
|
| 31 |
+
"authority": "<｜authority｜>",
|
| 32 |
+
"domain": "<｜domain｜>",
|
| 33 |
+
"title": "<｜title｜>",
|
| 34 |
+
"read_url": "<｜read_url｜>",
|
| 35 |
+
}
|
| 36 |
+
VALID_TASKS = set(DS_TASK_SP_TOKENS.keys())
|
| 37 |
+
|
| 38 |
+
# ============================================================
|
| 39 |
+
# Templates
|
| 40 |
+
# ============================================================
|
| 41 |
+
|
| 42 |
+
system_msg_template: str = "{content}"
|
| 43 |
+
user_msg_template: str = "{content}"
|
| 44 |
+
latest_reminder_msg_template: str = "{content}"
|
| 45 |
+
assistant_msg_template: str = "{reasoning}{content}{tool_calls}" + eos_token
|
| 46 |
+
assistant_msg_wo_eos_template: str = "{reasoning}{content}{tool_calls}"
|
| 47 |
+
thinking_template: str = "{reasoning_content}"
|
| 48 |
+
|
| 49 |
+
response_format_template: str = (
|
| 50 |
+
"## Response Format:\n\nYou MUST strictly adhere to the following schema to reply:\n{schema}"
|
| 51 |
+
)
|
| 52 |
+
tool_call_template: str = (
|
| 53 |
+
"<{dsml_token}invoke name=\"{name}\">\n{arguments}\n</{dsml_token}invoke>"
|
| 54 |
+
)
|
| 55 |
+
tool_calls_template = (
|
| 56 |
+
"<{dsml_token}{tc_block_name}>\n{tool_calls}\n</{dsml_token}{tc_block_name}>"
|
| 57 |
+
)
|
| 58 |
+
tool_calls_block_name: str = "tool_calls"
|
| 59 |
+
|
| 60 |
+
tool_output_template: str = (
|
| 61 |
+
"<tool_result>{content}</tool_result>"
|
| 62 |
+
)
|
| 63 |
+
|
| 64 |
+
REASONING_EFFORT_MAX = (
|
| 65 |
+
"Reasoning Effort: Absolute maximum with no shortcuts permitted.\n"
|
| 66 |
+
"You MUST be very thorough in your thinking and comprehensively decompose the problem to resolve the root cause, rigorously stress-testing your logic against all potential paths, edge cases, and adversarial scenarios.\n"
|
| 67 |
+
"Explicitly write out your entire deliberation process, documenting every intermediate step, considered alternative, and rejected hypothesis to ensure absolutely no assumption is left unchecked.\n\n"
|
| 68 |
+
)
|
| 69 |
+
|
| 70 |
+
TOOLS_TEMPLATE = """## Tools
|
| 71 |
+
|
| 72 |
+
You have access to a set of tools to help answer the user's question. You can invoke tools by writing a "<{dsml_token}tool_calls>" block like the following:
|
| 73 |
+
|
| 74 |
+
<{dsml_token}tool_calls>
|
| 75 |
+
<{dsml_token}invoke name="$TOOL_NAME">
|
| 76 |
+
<{dsml_token}parameter name="$PARAMETER_NAME" string="true|false">$PARAMETER_VALUE</{dsml_token}parameter>
|
| 77 |
+
...
|
| 78 |
+
</{dsml_token}invoke>
|
| 79 |
+
<{dsml_token}invoke name="$TOOL_NAME2">
|
| 80 |
+
...
|
| 81 |
+
</{dsml_token}invoke>
|
| 82 |
+
</{dsml_token}tool_calls>
|
| 83 |
+
|
| 84 |
+
String parameters should be specified as is and set `string="true"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string="false"`.
|
| 85 |
+
|
| 86 |
+
If thinking_mode is enabled (triggered by {thinking_start_token}), you MUST output your complete reasoning inside {thinking_start_token}...{thinking_end_token} BEFORE any tool calls or final response.
|
| 87 |
+
|
| 88 |
+
Otherwise, output directly after {thinking_end_token} with tool calls or final response.
|
| 89 |
+
|
| 90 |
+
### Available Tool Schemas
|
| 91 |
+
|
| 92 |
+
{tool_schemas}
|
| 93 |
+
|
| 94 |
+
You MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.
|
| 95 |
+
"""
|
| 96 |
+
|
| 97 |
+
# ============================================================
|
| 98 |
+
# Utility Functions
|
| 99 |
+
# ============================================================
|
| 100 |
+
|
| 101 |
+
def to_json(value: Any) -> str:
|
| 102 |
+
"""Serialize a value to JSON string."""
|
| 103 |
+
try:
|
| 104 |
+
return json.dumps(value, ensure_ascii=False)
|
| 105 |
+
except:
|
| 106 |
+
return json.dumps(value, ensure_ascii=True)
|
| 107 |
+
|
| 108 |
+
|
| 109 |
+
def tools_from_openai_format(tools):
|
| 110 |
+
"""Extract function definitions from OpenAI-format tool list."""
|
| 111 |
+
return [tool["function"] for tool in tools]
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
def tool_calls_from_openai_format(tool_calls):
|
| 115 |
+
"""Convert OpenAI-format tool calls to internal format."""
|
| 116 |
+
return [
|
| 117 |
+
{
|
| 118 |
+
"name": tool_call["function"]["name"],
|
| 119 |
+
"arguments": tool_call["function"]["arguments"],
|
| 120 |
+
}
|
| 121 |
+
for tool_call in tool_calls
|
| 122 |
+
]
|
| 123 |
+
|
| 124 |
+
|
| 125 |
+
def tool_calls_to_openai_format(tool_calls):
|
| 126 |
+
"""Convert internal tool calls to OpenAI format."""
|
| 127 |
+
return [
|
| 128 |
+
{
|
| 129 |
+
"type": "function",
|
| 130 |
+
"function": {
|
| 131 |
+
"name": tool_call["name"],
|
| 132 |
+
"arguments": tool_call["arguments"],
|
| 133 |
+
}
|
| 134 |
+
}
|
| 135 |
+
for tool_call in tool_calls
|
| 136 |
+
]
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
def encode_arguments_to_dsml(tool_call: Dict[str, str]) -> str:
|
| 140 |
+
"""
|
| 141 |
+
Encode tool call arguments into DSML parameter format.
|
| 142 |
+
|
| 143 |
+
Args:
|
| 144 |
+
tool_call: Dict with "name" and "arguments" (JSON string) keys.
|
| 145 |
+
|
| 146 |
+
Returns:
|
| 147 |
+
DSML-formatted parameter string.
|
| 148 |
+
"""
|
| 149 |
+
p_dsml_template = '<{dsml_token}parameter name="{key}" string="{is_str}">{value}</{dsml_token}parameter>'
|
| 150 |
+
P_dsml_strs = []
|
| 151 |
+
|
| 152 |
+
try:
|
| 153 |
+
arguments = json.loads(tool_call["arguments"])
|
| 154 |
+
except Exception as err:
|
| 155 |
+
arguments = {"arguments": tool_call["arguments"]}
|
| 156 |
+
|
| 157 |
+
for k, v in arguments.items():
|
| 158 |
+
p_dsml_str = p_dsml_template.format(
|
| 159 |
+
dsml_token=dsml_token,
|
| 160 |
+
key=k,
|
| 161 |
+
is_str="true" if isinstance(v, str) else "false",
|
| 162 |
+
value=v if isinstance(v, str) else to_json(v),
|
| 163 |
+
)
|
| 164 |
+
P_dsml_strs.append(p_dsml_str)
|
| 165 |
+
|
| 166 |
+
return "\n".join(P_dsml_strs)
|
| 167 |
+
|
| 168 |
+
|
| 169 |
+
def decode_dsml_to_arguments(tool_name: str, tool_args: Dict[str, Tuple[str, str]]) -> Dict[str, str]:
|
| 170 |
+
"""
|
| 171 |
+
Decode DSML parameters back to a tool call dict.
|
| 172 |
+
|
| 173 |
+
Args:
|
| 174 |
+
tool_name: Name of the tool.
|
| 175 |
+
tool_args: Dict mapping param_name -> (value, is_string_flag).
|
| 176 |
+
|
| 177 |
+
Returns:
|
| 178 |
+
Dict with "name" and "arguments" (JSON string) keys.
|
| 179 |
+
"""
|
| 180 |
+
def _decode_value(key: str, value: str, string: str):
|
| 181 |
+
if string == "true":
|
| 182 |
+
value = to_json(value)
|
| 183 |
+
return f"{to_json(key)}: {value}"
|
| 184 |
+
|
| 185 |
+
tool_args_json = "{" + ", ".join([_decode_value(k, v, string=is_str) for k, (v, is_str) in tool_args.items()]) + "}"
|
| 186 |
+
return dict(name=tool_name, arguments=tool_args_json)
|
| 187 |
+
|
| 188 |
+
|
| 189 |
+
def render_tools(tools: List[Dict[str, Union[str, Dict[str, Any]]]]) -> str:
|
| 190 |
+
"""
|
| 191 |
+
Render tool schemas into the system prompt format.
|
| 192 |
+
|
| 193 |
+
Args:
|
| 194 |
+
tools: List of tool schema dicts (each with name, description, parameters).
|
| 195 |
+
|
| 196 |
+
Returns:
|
| 197 |
+
Formatted tools section string.
|
| 198 |
+
"""
|
| 199 |
+
tools_json = [to_json(t) for t in tools]
|
| 200 |
+
|
| 201 |
+
return TOOLS_TEMPLATE.format(
|
| 202 |
+
tool_schemas="\n".join(tools_json),
|
| 203 |
+
dsml_token=dsml_token,
|
| 204 |
+
thinking_start_token=thinking_start_token,
|
| 205 |
+
thinking_end_token=thinking_end_token,
|
| 206 |
+
)
|
| 207 |
+
|
| 208 |
+
|
| 209 |
+
def find_last_user_index(messages: List[Dict[str, Any]]) -> int:
|
| 210 |
+
"""Find the index of the last user/developer message."""
|
| 211 |
+
last_user_index = -1
|
| 212 |
+
for idx in range(len(messages) - 1, -1, -1):
|
| 213 |
+
if messages[idx].get("role") in ["user", "developer"]:
|
| 214 |
+
last_user_index = idx
|
| 215 |
+
break
|
| 216 |
+
return last_user_index
|
| 217 |
+
|
| 218 |
+
|
| 219 |
+
# ============================================================
|
| 220 |
+
# Message Rendering
|
| 221 |
+
# ============================================================
|
| 222 |
+
|
| 223 |
+
def render_message(index: int, messages: List[Dict[str, Any]], thinking_mode: str, drop_thinking: bool = True, reasoning_effort: Optional[str] = None) -> str:
|
| 224 |
+
"""
|
| 225 |
+
Render a single message at the given index into its encoded string form.
|
| 226 |
+
|
| 227 |
+
This is the core function that converts each message in the conversation
|
| 228 |
+
into the DeepSeek-V4 format.
|
| 229 |
+
|
| 230 |
+
Args:
|
| 231 |
+
index: Index of the message to render.
|
| 232 |
+
messages: Full list of messages in the conversation.
|
| 233 |
+
thinking_mode: Either "chat" or "thinking".
|
| 234 |
+
drop_thinking: Whether to drop reasoning content from earlier turns.
|
| 235 |
+
reasoning_effort: Optional reasoning effort level ("max", "high", or None).
|
| 236 |
+
|
| 237 |
+
Returns:
|
| 238 |
+
Encoded string for this message.
|
| 239 |
+
"""
|
| 240 |
+
assert 0 <= index < len(messages)
|
| 241 |
+
assert thinking_mode in ["chat", "thinking"], f"Invalid thinking_mode `{thinking_mode}`"
|
| 242 |
+
|
| 243 |
+
prompt = ""
|
| 244 |
+
msg = messages[index]
|
| 245 |
+
last_user_idx = find_last_user_index(messages)
|
| 246 |
+
|
| 247 |
+
role = msg.get("role")
|
| 248 |
+
content = msg.get("content")
|
| 249 |
+
tools = msg.get("tools")
|
| 250 |
+
response_format = msg.get("response_format")
|
| 251 |
+
tool_calls = msg.get("tool_calls")
|
| 252 |
+
reasoning_content = msg.get("reasoning_content")
|
| 253 |
+
wo_eos = msg.get("wo_eos", False)
|
| 254 |
+
|
| 255 |
+
if tools:
|
| 256 |
+
tools = tools_from_openai_format(tools)
|
| 257 |
+
if tool_calls:
|
| 258 |
+
tool_calls = tool_calls_from_openai_format(tool_calls)
|
| 259 |
+
|
| 260 |
+
# Reasoning effort prefix (only at index 0 in thinking mode with max effort)
|
| 261 |
+
assert reasoning_effort in ['max', None, 'high'], f"Invalid reasoning effort: {reasoning_effort}"
|
| 262 |
+
if index == 0 and thinking_mode == "thinking" and reasoning_effort == 'max':
|
| 263 |
+
prompt += REASONING_EFFORT_MAX
|
| 264 |
+
|
| 265 |
+
if role == "system":
|
| 266 |
+
prompt += system_msg_template.format(content=content or "")
|
| 267 |
+
if tools:
|
| 268 |
+
prompt += "\n\n" + render_tools(tools)
|
| 269 |
+
if response_format:
|
| 270 |
+
prompt += "\n\n" + response_format_template.format(schema=to_json(response_format))
|
| 271 |
+
|
| 272 |
+
elif role == "developer":
|
| 273 |
+
assert content, f"Invalid message for role `{role}`: {msg}"
|
| 274 |
+
|
| 275 |
+
content_developer = USER_SP_TOKEN
|
| 276 |
+
content_developer += content
|
| 277 |
+
|
| 278 |
+
if tools:
|
| 279 |
+
content_developer += "\n\n" + render_tools(tools)
|
| 280 |
+
if response_format:
|
| 281 |
+
content_developer += "\n\n" + response_format_template.format(schema=to_json(response_format))
|
| 282 |
+
|
| 283 |
+
prompt += user_msg_template.format(content=content_developer)
|
| 284 |
+
|
| 285 |
+
elif role == "user":
|
| 286 |
+
prompt += USER_SP_TOKEN
|
| 287 |
+
|
| 288 |
+
# Handle content blocks (tool results mixed with text)
|
| 289 |
+
content_blocks = msg.get("content_blocks")
|
| 290 |
+
if content_blocks:
|
| 291 |
+
parts = []
|
| 292 |
+
for block in content_blocks:
|
| 293 |
+
block_type = block.get("type")
|
| 294 |
+
if block_type == "text":
|
| 295 |
+
parts.append(block.get("text", ""))
|
| 296 |
+
elif block_type == "tool_result":
|
| 297 |
+
tool_content = block.get("content", "")
|
| 298 |
+
if isinstance(tool_content, list):
|
| 299 |
+
text_parts = []
|
| 300 |
+
for b in tool_content:
|
| 301 |
+
if b.get("type") == "text":
|
| 302 |
+
text_parts.append(b.get("text", ""))
|
| 303 |
+
else:
|
| 304 |
+
text_parts.append(f"[Unsupported {b.get('type')}]")
|
| 305 |
+
tool_content = "\n\n".join(text_parts)
|
| 306 |
+
parts.append(tool_output_template.format(content=tool_content))
|
| 307 |
+
else:
|
| 308 |
+
parts.append(f"[Unsupported {block_type}]")
|
| 309 |
+
prompt += "\n\n".join(parts)
|
| 310 |
+
else:
|
| 311 |
+
prompt += content or ""
|
| 312 |
+
|
| 313 |
+
elif role == "latest_reminder":
|
| 314 |
+
prompt += LATEST_REMINDER_SP_TOKEN + latest_reminder_msg_template.format(content=content)
|
| 315 |
+
|
| 316 |
+
elif role == "tool":
|
| 317 |
+
raise NotImplementedError("deepseek_v4 merges tool messages into user; please preprocess with merge_tool_messages()")
|
| 318 |
+
|
| 319 |
+
elif role == "assistant":
|
| 320 |
+
thinking_part = ""
|
| 321 |
+
tc_content = ""
|
| 322 |
+
|
| 323 |
+
if tool_calls:
|
| 324 |
+
tc_list = [
|
| 325 |
+
tool_call_template.format(
|
| 326 |
+
dsml_token=dsml_token,
|
| 327 |
+
name=tc.get("name"),
|
| 328 |
+
arguments=encode_arguments_to_dsml(tc)
|
| 329 |
+
)
|
| 330 |
+
for tc in tool_calls
|
| 331 |
+
]
|
| 332 |
+
tc_content += '\n\n' + tool_calls_template.format(
|
| 333 |
+
dsml_token=dsml_token,
|
| 334 |
+
tool_calls="\n".join(tc_list),
|
| 335 |
+
tc_block_name=tool_calls_block_name,
|
| 336 |
+
)
|
| 337 |
+
|
| 338 |
+
summary_content = content or ""
|
| 339 |
+
rc = reasoning_content or ""
|
| 340 |
+
|
| 341 |
+
# Check if previous message has a task - if so, this is a task output (no thinking)
|
| 342 |
+
prev_has_task = index - 1 >= 0 and messages[index - 1].get("task") is not None
|
| 343 |
+
|
| 344 |
+
if thinking_mode == "thinking" and not prev_has_task:
|
| 345 |
+
if not drop_thinking or index > last_user_idx:
|
| 346 |
+
thinking_part = thinking_template.format(reasoning_content=rc) + thinking_end_token
|
| 347 |
+
else:
|
| 348 |
+
thinking_part = ""
|
| 349 |
+
|
| 350 |
+
if wo_eos:
|
| 351 |
+
prompt += assistant_msg_wo_eos_template.format(
|
| 352 |
+
reasoning=thinking_part,
|
| 353 |
+
content=summary_content,
|
| 354 |
+
tool_calls=tc_content,
|
| 355 |
+
)
|
| 356 |
+
else:
|
| 357 |
+
prompt += assistant_msg_template.format(
|
| 358 |
+
reasoning=thinking_part,
|
| 359 |
+
content=summary_content,
|
| 360 |
+
tool_calls=tc_content,
|
| 361 |
+
)
|
| 362 |
+
else:
|
| 363 |
+
raise NotImplementedError(f"Unknown role: {role}")
|
| 364 |
+
|
| 365 |
+
# Append transition tokens based on what follows
|
| 366 |
+
if index + 1 < len(messages) and messages[index + 1].get("role") not in ["assistant", "latest_reminder"]:
|
| 367 |
+
return prompt
|
| 368 |
+
|
| 369 |
+
task = messages[index].get("task")
|
| 370 |
+
if task is not None:
|
| 371 |
+
# Task special token for internal classification tasks
|
| 372 |
+
assert task in VALID_TASKS, f"Invalid task: '{task}'. Valid tasks are: {list(VALID_TASKS)}"
|
| 373 |
+
task_sp_token = DS_TASK_SP_TOKENS[task]
|
| 374 |
+
|
| 375 |
+
if task != "action":
|
| 376 |
+
# Non-action tasks: append task sp token directly after the message
|
| 377 |
+
prompt += task_sp_token
|
| 378 |
+
else:
|
| 379 |
+
# Action task: append Assistant + thinking token + action sp token
|
| 380 |
+
prompt += ASSISTANT_SP_TOKEN
|
| 381 |
+
prompt += thinking_end_token if thinking_mode != "thinking" else thinking_start_token
|
| 382 |
+
prompt += task_sp_token
|
| 383 |
+
|
| 384 |
+
elif messages[index].get("role") in ["user", "developer"]:
|
| 385 |
+
# Normal generation: append Assistant + thinking token
|
| 386 |
+
prompt += ASSISTANT_SP_TOKEN
|
| 387 |
+
if not drop_thinking and thinking_mode == "thinking":
|
| 388 |
+
prompt += thinking_start_token
|
| 389 |
+
elif drop_thinking and thinking_mode == "thinking" and index >= last_user_idx:
|
| 390 |
+
prompt += thinking_start_token
|
| 391 |
+
else:
|
| 392 |
+
prompt += thinking_end_token
|
| 393 |
+
|
| 394 |
+
return prompt
|
| 395 |
+
|
| 396 |
+
|
| 397 |
+
# ============================================================
|
| 398 |
+
# Preprocessing
|
| 399 |
+
# ============================================================
|
| 400 |
+
|
| 401 |
+
def merge_tool_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
|
| 402 |
+
"""
|
| 403 |
+
Merge tool messages into the preceding user message using content_blocks format.
|
| 404 |
+
|
| 405 |
+
DeepSeek-V4 does not have a standalone "tool" role; instead, tool results
|
| 406 |
+
are encoded as <tool_result> blocks within user messages.
|
| 407 |
+
|
| 408 |
+
This function converts a standard OpenAI-format conversation (with separate
|
| 409 |
+
"tool" role messages) into V4 format where tool results are merged into
|
| 410 |
+
user messages.
|
| 411 |
+
|
| 412 |
+
Args:
|
| 413 |
+
messages: List of message dicts in OpenAI format.
|
| 414 |
+
|
| 415 |
+
Returns:
|
| 416 |
+
Processed message list with tool messages merged into user messages.
|
| 417 |
+
"""
|
| 418 |
+
merged: List[Dict[str, Any]] = []
|
| 419 |
+
|
| 420 |
+
for msg in messages:
|
| 421 |
+
msg = copy.deepcopy(msg)
|
| 422 |
+
role = msg.get("role")
|
| 423 |
+
|
| 424 |
+
if role == "tool":
|
| 425 |
+
# Convert tool message to a user message with tool_result block
|
| 426 |
+
tool_block = {
|
| 427 |
+
"type": "tool_result",
|
| 428 |
+
"tool_use_id": msg.get("tool_call_id", ""),
|
| 429 |
+
"content": msg.get("content", ""),
|
| 430 |
+
}
|
| 431 |
+
# Merge into previous message if it's already a user (merged tool)
|
| 432 |
+
if merged and merged[-1].get("role") == "user" and "content_blocks" in merged[-1]:
|
| 433 |
+
merged[-1]["content_blocks"].append(tool_block)
|
| 434 |
+
else:
|
| 435 |
+
merged.append({
|
| 436 |
+
"role": "user",
|
| 437 |
+
"content_blocks": [tool_block],
|
| 438 |
+
})
|
| 439 |
+
elif role == "user":
|
| 440 |
+
text_block = {"type": "text", "text": msg.get("content", "")}
|
| 441 |
+
if merged and merged[-1].get("role") == "user" and "content_blocks" in merged[-1] and merged[-1].get("task") is None:
|
| 442 |
+
merged[-1]["content_blocks"].append(text_block)
|
| 443 |
+
else:
|
| 444 |
+
new_msg = {
|
| 445 |
+
"role": "user",
|
| 446 |
+
"content": msg.get("content", ""),
|
| 447 |
+
"content_blocks": [text_block],
|
| 448 |
+
}
|
| 449 |
+
# Preserve extra fields (task, wo_eos, mask, etc.)
|
| 450 |
+
for key in ("task", "wo_eos", "mask"):
|
| 451 |
+
if key in msg:
|
| 452 |
+
new_msg[key] = msg[key]
|
| 453 |
+
merged.append(new_msg)
|
| 454 |
+
else:
|
| 455 |
+
merged.append(msg)
|
| 456 |
+
|
| 457 |
+
return merged
|
| 458 |
+
|
| 459 |
+
|
| 460 |
+
def sort_tool_results_by_call_order(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
|
| 461 |
+
"""
|
| 462 |
+
Sort tool_result blocks within user messages by the order of tool_calls
|
| 463 |
+
in the preceding assistant message.
|
| 464 |
+
|
| 465 |
+
Args:
|
| 466 |
+
messages: Preprocessed message list (after merge_tool_messages).
|
| 467 |
+
|
| 468 |
+
Returns:
|
| 469 |
+
Message list with sorted tool result blocks.
|
| 470 |
+
"""
|
| 471 |
+
last_tool_call_order: Dict[str, int] = {}
|
| 472 |
+
|
| 473 |
+
for msg in messages:
|
| 474 |
+
role = msg.get("role")
|
| 475 |
+
if role == "assistant" and msg.get("tool_calls"):
|
| 476 |
+
last_tool_call_order = {}
|
| 477 |
+
for idx, tc in enumerate(msg["tool_calls"]):
|
| 478 |
+
tc_id = tc.get("id") or tc.get("function", {}).get("id", "")
|
| 479 |
+
if tc_id:
|
| 480 |
+
last_tool_call_order[tc_id] = idx
|
| 481 |
+
|
| 482 |
+
elif role == "user" and msg.get("content_blocks"):
|
| 483 |
+
tool_blocks = [b for b in msg["content_blocks"] if b.get("type") == "tool_result"]
|
| 484 |
+
if len(tool_blocks) > 1 and last_tool_call_order:
|
| 485 |
+
sorted_blocks = sorted(
|
| 486 |
+
tool_blocks,
|
| 487 |
+
key=lambda b: last_tool_call_order.get(b.get("tool_use_id", ""), 0)
|
| 488 |
+
)
|
| 489 |
+
sorted_idx = 0
|
| 490 |
+
new_blocks = []
|
| 491 |
+
for block in msg["content_blocks"]:
|
| 492 |
+
if block.get("type") == "tool_result":
|
| 493 |
+
new_blocks.append(sorted_blocks[sorted_idx])
|
| 494 |
+
sorted_idx += 1
|
| 495 |
+
else:
|
| 496 |
+
new_blocks.append(block)
|
| 497 |
+
msg["content_blocks"] = new_blocks
|
| 498 |
+
|
| 499 |
+
return messages
|
| 500 |
+
|
| 501 |
+
|
| 502 |
+
# ============================================================
|
| 503 |
+
# Main Encoding Function
|
| 504 |
+
# ============================================================
|
| 505 |
+
|
| 506 |
+
def encode_messages(
|
| 507 |
+
messages: List[Dict[str, Any]],
|
| 508 |
+
thinking_mode: str,
|
| 509 |
+
context: Optional[List[Dict[str, Any]]] = None,
|
| 510 |
+
drop_thinking: bool = True,
|
| 511 |
+
add_default_bos_token: bool = True,
|
| 512 |
+
reasoning_effort: Optional[str] = None,
|
| 513 |
+
) -> str:
|
| 514 |
+
"""
|
| 515 |
+
Encode a list of messages into the DeepSeek-V4 prompt format.
|
| 516 |
+
|
| 517 |
+
This is the main entry point for encoding conversations. It handles:
|
| 518 |
+
- BOS token insertion
|
| 519 |
+
- Thinking mode with optional reasoning content dropping
|
| 520 |
+
- Tool message merging into user messages
|
| 521 |
+
- Multi-turn conversation context
|
| 522 |
+
|
| 523 |
+
Args:
|
| 524 |
+
messages: List of message dicts to encode.
|
| 525 |
+
thinking_mode: Either "chat" or "thinking".
|
| 526 |
+
context: Optional preceding context messages (already encoded prefix).
|
| 527 |
+
drop_thinking: If True, drop reasoning_content from earlier assistant turns
|
| 528 |
+
(only keep reasoning for messages after the last user message).
|
| 529 |
+
add_default_bos_token: Whether to prepend BOS token at conversation start.
|
| 530 |
+
reasoning_effort: Optional reasoning effort level ("max", "high", or None).
|
| 531 |
+
|
| 532 |
+
Returns:
|
| 533 |
+
The encoded prompt string.
|
| 534 |
+
"""
|
| 535 |
+
context = context if context else []
|
| 536 |
+
|
| 537 |
+
# Preprocess: merge tool messages and sort tool results
|
| 538 |
+
messages = merge_tool_messages(messages)
|
| 539 |
+
messages = sort_tool_results_by_call_order(context + messages)[len(context):]
|
| 540 |
+
if context:
|
| 541 |
+
context = merge_tool_messages(context)
|
| 542 |
+
context = sort_tool_results_by_call_order(context)
|
| 543 |
+
|
| 544 |
+
full_messages = context + messages
|
| 545 |
+
|
| 546 |
+
prompt = bos_token if add_default_bos_token and len(context) == 0 else ""
|
| 547 |
+
|
| 548 |
+
# Resolve drop_thinking: if any message has tools defined, don't drop thinking
|
| 549 |
+
effective_drop_thinking = drop_thinking
|
| 550 |
+
if any(m.get("tools") for m in full_messages):
|
| 551 |
+
effective_drop_thinking = False
|
| 552 |
+
|
| 553 |
+
if thinking_mode == "thinking" and effective_drop_thinking:
|
| 554 |
+
full_messages = _drop_thinking_messages(full_messages)
|
| 555 |
+
# After dropping, recalculate how many messages to render
|
| 556 |
+
# (context may have shrunk too)
|
| 557 |
+
num_to_render = len(full_messages) - len(_drop_thinking_messages(context))
|
| 558 |
+
context_len = len(full_messages) - num_to_render
|
| 559 |
+
else:
|
| 560 |
+
num_to_render = len(messages)
|
| 561 |
+
context_len = len(context)
|
| 562 |
+
|
| 563 |
+
for idx in range(num_to_render):
|
| 564 |
+
prompt += render_message(
|
| 565 |
+
idx + context_len,
|
| 566 |
+
full_messages,
|
| 567 |
+
thinking_mode=thinking_mode,
|
| 568 |
+
drop_thinking=effective_drop_thinking,
|
| 569 |
+
reasoning_effort=reasoning_effort,
|
| 570 |
+
)
|
| 571 |
+
|
| 572 |
+
return prompt
|
| 573 |
+
|
| 574 |
+
|
| 575 |
+
def _drop_thinking_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
|
| 576 |
+
"""
|
| 577 |
+
Drop reasoning_content and non-essential messages before the last user message.
|
| 578 |
+
|
| 579 |
+
Behavior:
|
| 580 |
+
- Messages with role in ["user", "system", "tool", "latest_reminder"] are always kept.
|
| 581 |
+
- Messages at or after the last user index are always kept.
|
| 582 |
+
- Assistant messages before the last user get reasoning_content removed.
|
| 583 |
+
- Developer messages before the last user are dropped entirely.
|
| 584 |
+
"""
|
| 585 |
+
last_user_idx = find_last_user_index(messages)
|
| 586 |
+
result = []
|
| 587 |
+
keep_roles = {"user", "system", "tool", "latest_reminder", "direct_search_results"}
|
| 588 |
+
|
| 589 |
+
for idx, msg in enumerate(messages):
|
| 590 |
+
role = msg.get("role")
|
| 591 |
+
if role in keep_roles or idx >= last_user_idx:
|
| 592 |
+
result.append(msg)
|
| 593 |
+
elif role == "assistant":
|
| 594 |
+
msg = copy.copy(msg)
|
| 595 |
+
msg.pop("reasoning_content", None)
|
| 596 |
+
result.append(msg)
|
| 597 |
+
# developer and other roles before last_user_idx are dropped
|
| 598 |
+
|
| 599 |
+
return result
|
| 600 |
+
|
| 601 |
+
|
| 602 |
+
# ============================================================
|
| 603 |
+
# Parsing (Decoding model output)
|
| 604 |
+
# ============================================================
|
| 605 |
+
|
| 606 |
+
def _read_until_stop(index: int, text: str, stop: List[str]) -> Tuple[int, str, Optional[str]]:
|
| 607 |
+
"""
|
| 608 |
+
Read text from index until one of the stop strings is found.
|
| 609 |
+
|
| 610 |
+
Returns:
|
| 611 |
+
Tuple of (new_index, content_before_stop, matched_stop_string_or_None).
|
| 612 |
+
"""
|
| 613 |
+
min_pos = len(text)
|
| 614 |
+
matched_stop = None
|
| 615 |
+
|
| 616 |
+
for s in stop:
|
| 617 |
+
pos = text.find(s, index)
|
| 618 |
+
if pos != -1 and pos < min_pos:
|
| 619 |
+
min_pos = pos
|
| 620 |
+
matched_stop = s
|
| 621 |
+
|
| 622 |
+
if matched_stop:
|
| 623 |
+
content = text[index:min_pos]
|
| 624 |
+
return min_pos + len(matched_stop), content, matched_stop
|
| 625 |
+
else:
|
| 626 |
+
content = text[index:]
|
| 627 |
+
return len(text), content, None
|
| 628 |
+
|
| 629 |
+
|
| 630 |
+
def parse_tool_calls(index: int, text: str) -> Tuple[int, Optional[str], List[Dict[str, str]]]:
|
| 631 |
+
"""
|
| 632 |
+
Parse DSML tool calls from text starting at the given index.
|
| 633 |
+
|
| 634 |
+
Args:
|
| 635 |
+
index: Starting position in text.
|
| 636 |
+
text: The full text to parse.
|
| 637 |
+
|
| 638 |
+
Returns:
|
| 639 |
+
Tuple of (new_index, last_stop_token, list_of_tool_call_dicts).
|
| 640 |
+
Each tool call dict has "name" and "arguments" keys.
|
| 641 |
+
"""
|
| 642 |
+
tool_calls: List[Dict[str, Any]] = []
|
| 643 |
+
stop_token = None
|
| 644 |
+
tool_calls_end_token = f"</{dsml_token}{tool_calls_block_name}>"
|
| 645 |
+
|
| 646 |
+
while index < len(text):
|
| 647 |
+
index, _, stop_token = _read_until_stop(index, text, [f"<{dsml_token}invoke", tool_calls_end_token])
|
| 648 |
+
if _ != ">\n":
|
| 649 |
+
raise ValueError(f"Tool call format error: expected '>\\n' but got '{_}'")
|
| 650 |
+
|
| 651 |
+
if stop_token == tool_calls_end_token:
|
| 652 |
+
break
|
| 653 |
+
|
| 654 |
+
if stop_token is None:
|
| 655 |
+
raise ValueError("Missing special token in tool calls")
|
| 656 |
+
|
| 657 |
+
index, tool_name_content, stop_token = _read_until_stop(index, text, [f"<{dsml_token}parameter", f"</{dsml_token}invoke"])
|
| 658 |
+
|
| 659 |
+
p_tool_name = re.findall(r'^\s*name="(.*?)">\n$', tool_name_content, flags=re.DOTALL)
|
| 660 |
+
if len(p_tool_name) != 1:
|
| 661 |
+
raise ValueError(f"Tool name format error: '{tool_name_content}'")
|
| 662 |
+
tool_name = p_tool_name[0]
|
| 663 |
+
|
| 664 |
+
tool_args: Dict[str, Tuple[str, str]] = {}
|
| 665 |
+
while stop_token == f"<{dsml_token}parameter":
|
| 666 |
+
index, param_content, stop_token = _read_until_stop(index, text, [f"/{dsml_token}parameter"])
|
| 667 |
+
|
| 668 |
+
param_kv = re.findall(r'^ name="(.*?)" string="(true|false)">(.*?)<$', param_content, flags=re.DOTALL)
|
| 669 |
+
if len(param_kv) != 1:
|
| 670 |
+
raise ValueError(f"Parameter format error: '{param_content}'")
|
| 671 |
+
param_name, string, param_value = param_kv[0]
|
| 672 |
+
|
| 673 |
+
if param_name in tool_args:
|
| 674 |
+
raise ValueError(f"Duplicate parameter name: '{param_name}'")
|
| 675 |
+
tool_args[param_name] = (param_value, string)
|
| 676 |
+
|
| 677 |
+
index, content, stop_token = _read_until_stop(index, text, [f"<{dsml_token}parameter", f"</{dsml_token}invoke"])
|
| 678 |
+
if content != ">\n":
|
| 679 |
+
raise ValueError(f"Parameter format error: expected '>\\n' but got '{content}'")
|
| 680 |
+
|
| 681 |
+
tool_call = decode_dsml_to_arguments(tool_name=tool_name, tool_args=tool_args)
|
| 682 |
+
tool_calls.append(tool_call)
|
| 683 |
+
|
| 684 |
+
return index, stop_token, tool_calls
|
| 685 |
+
|
| 686 |
+
|
| 687 |
+
def parse_message_from_completion_text(text: str, thinking_mode: str) -> Dict[str, Any]:
|
| 688 |
+
"""
|
| 689 |
+
Parse a model completion text into a structured assistant message.
|
| 690 |
+
|
| 691 |
+
This function takes the raw text output from the model (a single assistant turn)
|
| 692 |
+
and extracts:
|
| 693 |
+
- reasoning_content (thinking block)
|
| 694 |
+
- content (summary/response)
|
| 695 |
+
- tool_calls (if any)
|
| 696 |
+
|
| 697 |
+
NOTE: This function is designed to parse only correctly formatted strings and
|
| 698 |
+
will raise ValueError for malformed output.
|
| 699 |
+
|
| 700 |
+
Args:
|
| 701 |
+
text: The raw completion text (including EOS token).
|
| 702 |
+
thinking_mode: Either "chat" or "thinking".
|
| 703 |
+
|
| 704 |
+
Returns:
|
| 705 |
+
Dict with keys: "role", "content", "reasoning_content", "tool_calls".
|
| 706 |
+
tool_calls are in OpenAI format.
|
| 707 |
+
"""
|
| 708 |
+
summary_content, reasoning_content, tool_calls = "", "", []
|
| 709 |
+
index, stop_token = 0, None
|
| 710 |
+
tool_calls_start_token = f"\n\n<{dsml_token}{tool_calls_block_name}"
|
| 711 |
+
|
| 712 |
+
is_thinking = thinking_mode == "thinking"
|
| 713 |
+
is_tool_calling = False
|
| 714 |
+
|
| 715 |
+
if is_thinking:
|
| 716 |
+
index, content_delta, stop_token = _read_until_stop(index, text, [thinking_end_token, tool_calls_start_token])
|
| 717 |
+
reasoning_content = content_delta
|
| 718 |
+
assert stop_token == thinking_end_token, "Invalid thinking format: missing </think>"
|
| 719 |
+
|
| 720 |
+
index, content_delta, stop_token = _read_until_stop(index, text, [eos_token, tool_calls_start_token])
|
| 721 |
+
summary_content = content_delta
|
| 722 |
+
if stop_token == tool_calls_start_token:
|
| 723 |
+
is_tool_calling = True
|
| 724 |
+
else:
|
| 725 |
+
assert stop_token == eos_token, "Invalid format: missing EOS token"
|
| 726 |
+
|
| 727 |
+
if is_tool_calling:
|
| 728 |
+
index, stop_token, tool_calls = parse_tool_calls(index, text)
|
| 729 |
+
|
| 730 |
+
index, tool_ends_text, stop_token = _read_until_stop(index, text, [eos_token])
|
| 731 |
+
assert not tool_ends_text, "Unexpected content after tool calls"
|
| 732 |
+
|
| 733 |
+
assert len(text) == index and stop_token in [eos_token, None], "Unexpected content at end"
|
| 734 |
+
|
| 735 |
+
for sp_token in [bos_token, eos_token, thinking_start_token, thinking_end_token, dsml_token]:
|
| 736 |
+
assert sp_token not in summary_content and sp_token not in reasoning_content, \
|
| 737 |
+
f"Unexpected special token '{sp_token}' in content"
|
| 738 |
+
|
| 739 |
+
return {
|
| 740 |
+
"role": "assistant",
|
| 741 |
+
"content": summary_content,
|
| 742 |
+
"reasoning_content": reasoning_content,
|
| 743 |
+
"tool_calls": tool_calls_to_openai_format(tool_calls)
|
| 744 |
+
}
|
encoding/test_encoding_dsv4.py
ADDED
|
@@ -0,0 +1,89 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Test suite for DeepSeek-V4 Encoding.
|
| 3 |
+
|
| 4 |
+
Run: python test_encoding_dsv4.py
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
import json
|
| 8 |
+
import os
|
| 9 |
+
|
| 10 |
+
from encoding_dsv4 import encode_messages, parse_message_from_completion_text
|
| 11 |
+
|
| 12 |
+
TESTS_DIR = os.path.join(os.path.dirname(__file__), "tests")
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
def test_case_1():
|
| 16 |
+
"""Thinking mode with tool calls (multi-turn, tool results merged into user)."""
|
| 17 |
+
with open(os.path.join(TESTS_DIR, "test_input_1.json")) as f:
|
| 18 |
+
td = json.load(f)
|
| 19 |
+
messages = td["messages"]
|
| 20 |
+
messages[0]["tools"] = td["tools"]
|
| 21 |
+
gold = open(os.path.join(TESTS_DIR, "test_output_1.txt")).read()
|
| 22 |
+
prompt = encode_messages(messages, thinking_mode="thinking")
|
| 23 |
+
assert prompt == gold
|
| 24 |
+
|
| 25 |
+
# Parse: assistant turn with tool call
|
| 26 |
+
marker = "<｜Assistant｜><think>"
|
| 27 |
+
first_start = prompt.find(marker) + len(marker)
|
| 28 |
+
first_end = prompt.find("<｜User｜>", first_start)
|
| 29 |
+
parsed_tc = parse_message_from_completion_text(prompt[first_start:first_end], thinking_mode="thinking")
|
| 30 |
+
assert parsed_tc["reasoning_content"] == "The user wants to know the weather in Beijing. I should use the get_weather tool."
|
| 31 |
+
assert parsed_tc["content"] == ""
|
| 32 |
+
assert len(parsed_tc["tool_calls"]) == 1
|
| 33 |
+
assert parsed_tc["tool_calls"][0]["function"]["name"] == "get_weather"
|
| 34 |
+
assert json.loads(parsed_tc["tool_calls"][0]["function"]["arguments"]) == {"location": "Beijing", "unit": "celsius"}
|
| 35 |
+
|
| 36 |
+
# Parse: final assistant turn with content
|
| 37 |
+
last_start = prompt.rfind(marker) + len(marker)
|
| 38 |
+
parsed_final = parse_message_from_completion_text(prompt[last_start:], thinking_mode="thinking")
|
| 39 |
+
assert parsed_final["reasoning_content"] == "Got the weather data. Let me format a nice response."
|
| 40 |
+
assert "22°C" in parsed_final["content"]
|
| 41 |
+
assert parsed_final["tool_calls"] == []
|
| 42 |
+
|
| 43 |
+
print(" [PASS] case 1: thinking with tools (encode + parse)")
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
def test_case_2():
|
| 47 |
+
"""Thinking mode without tools (drop_thinking removes earlier reasoning)."""
|
| 48 |
+
messages = json.load(open(os.path.join(TESTS_DIR, "test_input_2.json")))
|
| 49 |
+
gold = open(os.path.join(TESTS_DIR, "test_output_2.txt")).read()
|
| 50 |
+
prompt = encode_messages(messages, thinking_mode="thinking")
|
| 51 |
+
assert prompt == gold
|
| 52 |
+
|
| 53 |
+
# Parse: last assistant turn
|
| 54 |
+
marker = "<｜Assistant｜><think>"
|
| 55 |
+
last_start = prompt.rfind(marker) + len(marker)
|
| 56 |
+
parsed = parse_message_from_completion_text(prompt[last_start:], thinking_mode="thinking")
|
| 57 |
+
assert parsed["reasoning_content"] == "The user asks about the capital of France. It is Paris."
|
| 58 |
+
assert parsed["content"] == "The capital of France is Paris."
|
| 59 |
+
assert parsed["tool_calls"] == []
|
| 60 |
+
|
| 61 |
+
# Verify drop_thinking: first assistant's reasoning should be absent
|
| 62 |
+
assert "The user said hello" not in prompt
|
| 63 |
+
|
| 64 |
+
print(" [PASS] case 2: thinking without tools (encode + parse)")
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
def test_case_3():
|
| 68 |
+
"""Interleaved thinking + search (developer with tools, latest_reminder)."""
|
| 69 |
+
messages = json.load(open(os.path.join(TESTS_DIR, "test_input_3.json")))
|
| 70 |
+
gold = open(os.path.join(TESTS_DIR, "test_output_3.txt")).read()
|
| 71 |
+
assert encode_messages(messages, thinking_mode="thinking") == gold
|
| 72 |
+
print(" [PASS] case 3: interleaved thinking + search")
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
def test_case_4():
|
| 76 |
+
"""Quick instruction task with latest_reminder (chat mode, action task)."""
|
| 77 |
+
messages = json.load(open(os.path.join(TESTS_DIR, "test_input_4.json")))
|
| 78 |
+
gold = open(os.path.join(TESTS_DIR, "test_output_4.txt")).read()
|
| 79 |
+
assert encode_messages(messages, thinking_mode="chat") == gold
|
| 80 |
+
print(" [PASS] case 4: quick instruction task")
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
if __name__ == "__main__":
|
| 84 |
+
print("Running DeepSeek-V4 Encoding Tests...\n")
|
| 85 |
+
test_case_1()
|
| 86 |
+
test_case_2()
|
| 87 |
+
test_case_3()
|
| 88 |
+
test_case_4()
|
| 89 |
+
print("\nAll 4 tests passed!")
|
encoding/tests/test_input_1.json
ADDED
|
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"tools": [
|
| 3 |
+
{
|
| 4 |
+
"type": "function",
|
| 5 |
+
"function": {
|
| 6 |
+
"name": "get_weather",
|
| 7 |
+
"description": "Get the weather for a specific location",
|
| 8 |
+
"parameters": {
|
| 9 |
+
"type": "object",
|
| 10 |
+
"properties": {
|
| 11 |
+
"location": {
|
| 12 |
+
"type": "string",
|
| 13 |
+
"description": "The city name"
|
| 14 |
+
},
|
| 15 |
+
"unit": {
|
| 16 |
+
"type": "string",
|
| 17 |
+
"enum": ["celsius", "fahrenheit"],
|
| 18 |
+
"description": "Temperature unit"
|
| 19 |
+
}
|
| 20 |
+
},
|
| 21 |
+
"required": ["location"]
|
| 22 |
+
}
|
| 23 |
+
}
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "function",
|
| 27 |
+
"function": {
|
| 28 |
+
"name": "search",
|
| 29 |
+
"description": "Search the web for information",
|
| 30 |
+
"parameters": {
|
| 31 |
+
"type": "object",
|
| 32 |
+
"properties": {
|
| 33 |
+
"query": {
|
| 34 |
+
"type": "string",
|
| 35 |
+
"description": "Search query"
|
| 36 |
+
},
|
| 37 |
+
"num_results": {
|
| 38 |
+
"type": "integer",
|
| 39 |
+
"description": "Number of results to return"
|
| 40 |
+
}
|
| 41 |
+
},
|
| 42 |
+
"required": ["query"]
|
| 43 |
+
}
|
| 44 |
+
}
|
| 45 |
+
}
|
| 46 |
+
],
|
| 47 |
+
"messages": [
|
| 48 |
+
{
|
| 49 |
+
"role": "system",
|
| 50 |
+
"content": "You are a helpful assistant."
|
| 51 |
+
},
|
| 52 |
+
{
|
| 53 |
+
"role": "user",
|
| 54 |
+
"content": "What's the weather in Beijing?"
|
| 55 |
+
},
|
| 56 |
+
{
|
| 57 |
+
"role": "assistant",
|
| 58 |
+
"reasoning_content": "The user wants to know the weather in Beijing. I should use the get_weather tool.",
|
| 59 |
+
"tool_calls": [
|
| 60 |
+
{
|
| 61 |
+
"id": "call_001",
|
| 62 |
+
"type": "function",
|
| 63 |
+
"function": {
|
| 64 |
+
"name": "get_weather",
|
| 65 |
+
"arguments": "{\"location\": \"Beijing\", \"unit\": \"celsius\"}"
|
| 66 |
+
}
|
| 67 |
+
}
|
| 68 |
+
]
|
| 69 |
+
},
|
| 70 |
+
{
|
| 71 |
+
"role": "tool",
|
| 72 |
+
"tool_call_id": "call_001",
|
| 73 |
+
"content": "{\"temperature\": 22, \"condition\": \"sunny\", \"humidity\": 45}"
|
| 74 |
+
},
|
| 75 |
+
{
|
| 76 |
+
"role": "assistant",
|
| 77 |
+
"reasoning_content": "Got the weather data. Let me format a nice response.",
|
| 78 |
+
"content": "The weather in Beijing is currently sunny with a temperature of 22°C and 45% humidity."
|
| 79 |
+
}
|
| 80 |
+
]
|
| 81 |
+
}
|
encoding/tests/test_input_2.json
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"role": "system",
|
| 4 |
+
"content": "You are a helpful assistant."
|
| 5 |
+
},
|
| 6 |
+
{
|
| 7 |
+
"role": "user",
|
| 8 |
+
"content": "Hello"
|
| 9 |
+
},
|
| 10 |
+
{
|
| 11 |
+
"role": "assistant",
|
| 12 |
+
"reasoning_content": "The user said hello, I should greet back.",
|
| 13 |
+
"content": "Hi there! How can I help you?"
|
| 14 |
+
},
|
| 15 |
+
{
|
| 16 |
+
"role": "user",
|
| 17 |
+
"content": "What is the capital of France?"
|
| 18 |
+
},
|
| 19 |
+
{
|
| 20 |
+
"role": "assistant",
|
| 21 |
+
"reasoning_content": "The user asks about the capital of France. It is Paris.",
|
| 22 |
+
"content": "The capital of France is Paris."
|
| 23 |
+
}
|
| 24 |
+
]
|
encoding/tests/test_input_3.json
ADDED
|
@@ -0,0 +1,159 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"role": "system",
|
| 4 |
+
"content": "该助手为DeepSeek，由深度求索公司创造。"
|
| 5 |
+
},
|
| 6 |
+
{
|
| 7 |
+
"role": "latest_reminder",
|
| 8 |
+
"content": "2026-02-21,星期六,广州,App,中文"
|
| 9 |
+
},
|
| 10 |
+
{
|
| 11 |
+
"role": "developer",
|
| 12 |
+
"content": "小柴胡冲剂和布洛芬能一起吃吗？\n\nCITATION FORMAT: 【{cursor_id}†L{start_line_id}(-L{end_line_id})?】",
|
| 13 |
+
"tools": [
|
| 14 |
+
{
|
| 15 |
+
"type": "function",
|
| 16 |
+
"function": {
|
| 17 |
+
"name": "search",
|
| 18 |
+
"description": "Web search. Split multiple queries with '||'.",
|
| 19 |
+
"parameters": {
|
| 20 |
+
"type": "object",
|
| 21 |
+
"properties": {
|
| 22 |
+
"queries": {
|
| 23 |
+
"type": "string",
|
| 24 |
+
"description": "query1||query2"
|
| 25 |
+
}
|
| 26 |
+
},
|
| 27 |
+
"required": [
|
| 28 |
+
"queries"
|
| 29 |
+
],
|
| 30 |
+
"additionalProperties": false,
|
| 31 |
+
"$schema": "http://json-schema.org/draft-07/schema#"
|
| 32 |
+
}
|
| 33 |
+
}
|
| 34 |
+
},
|
| 35 |
+
{
|
| 36 |
+
"type": "function",
|
| 37 |
+
"function": {
|
| 38 |
+
"name": "open",
|
| 39 |
+
"description": "Batch open IDs (format 【{id}†...】) or URLs.",
|
| 40 |
+
"parameters": {
|
| 41 |
+
"type": "object",
|
| 42 |
+
"properties": {
|
| 43 |
+
"open_list": {
|
| 44 |
+
"type": "array",
|
| 45 |
+
"items": {
|
| 46 |
+
"type": "object",
|
| 47 |
+
"properties": {
|
| 48 |
+
"id": {
|
| 49 |
+
"description": "ID or URL",
|
| 50 |
+
"anyOf": [
|
| 51 |
+
{
|
| 52 |
+
"type": "integer"
|
| 53 |
+
},
|
| 54 |
+
{
|
| 55 |
+
"type": "string"
|
| 56 |
+
}
|
| 57 |
+
],
|
| 58 |
+
"default": -1
|
| 59 |
+
},
|
| 60 |
+
"cursor": {
|
| 61 |
+
"type": "integer",
|
| 62 |
+
"description": "",
|
| 63 |
+
"default": -1
|
| 64 |
+
},
|
| 65 |
+
"loc": {
|
| 66 |
+
"type": "integer",
|
| 67 |
+
"description": "Start line",
|
| 68 |
+
"default": -1
|
| 69 |
+
},
|
| 70 |
+
"num_lines": {
|
| 71 |
+
"type": "integer",
|
| 72 |
+
"description": "",
|
| 73 |
+
"default": -1
|
| 74 |
+
},
|
| 75 |
+
"view_source": {
|
| 76 |
+
"type": "boolean",
|
| 77 |
+
"description": "",
|
| 78 |
+
"default": false
|
| 79 |
+
}
|
| 80 |
+
},
|
| 81 |
+
"additionalProperties": false
|
| 82 |
+
},
|
| 83 |
+
"description": ""
|
| 84 |
+
}
|
| 85 |
+
},
|
| 86 |
+
"required": [
|
| 87 |
+
"open_list"
|
| 88 |
+
],
|
| 89 |
+
"additionalProperties": false,
|
| 90 |
+
"$schema": "http://json-schema.org/draft-07/schema#"
|
| 91 |
+
}
|
| 92 |
+
}
|
| 93 |
+
},
|
| 94 |
+
{
|
| 95 |
+
"type": "function",
|
| 96 |
+
"function": {
|
| 97 |
+
"name": "find",
|
| 98 |
+
"description": "Find exact text pattern in pages.",
|
| 99 |
+
"parameters": {
|
| 100 |
+
"type": "object",
|
| 101 |
+
"properties": {
|
| 102 |
+
"find_list": {
|
| 103 |
+
"type": "array",
|
| 104 |
+
"items": {
|
| 105 |
+
"type": "object",
|
| 106 |
+
"properties": {
|
| 107 |
+
"pattern": {
|
| 108 |
+
"type": "string",
|
| 109 |
+
"description": ""
|
| 110 |
+
},
|
| 111 |
+
"cursor": {
|
| 112 |
+
"type": "integer",
|
| 113 |
+
"description": "",
|
| 114 |
+
"default": -1
|
| 115 |
+
}
|
| 116 |
+
},
|
| 117 |
+
"required": [
|
| 118 |
+
"pattern"
|
| 119 |
+
],
|
| 120 |
+
"additionalProperties": false
|
| 121 |
+
},
|
| 122 |
+
"description": ""
|
| 123 |
+
}
|
| 124 |
+
},
|
| 125 |
+
"required": [
|
| 126 |
+
"find_list"
|
| 127 |
+
],
|
| 128 |
+
"additionalProperties": false,
|
| 129 |
+
"$schema": "http://json-schema.org/draft-07/schema#"
|
| 130 |
+
}
|
| 131 |
+
}
|
| 132 |
+
}
|
| 133 |
+
]
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"role": "assistant",
|
| 137 |
+
"content": "",
|
| 138 |
+
"reasoning_content": "用户想知道小柴胡冲剂和布洛芬能否一起服用。",
|
| 139 |
+
"tool_calls": [
|
| 140 |
+
{
|
| 141 |
+
"type": "function",
|
| 142 |
+
"function": {
|
| 143 |
+
"name": "search",
|
| 144 |
+
"arguments": "{\"queries\": \"小柴胡冲剂 布洛芬 相互作用 一起吃\"}"
|
| 145 |
+
}
|
| 146 |
+
}
|
| 147 |
+
]
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"role": "tool",
|
| 151 |
+
"content": "[0]"
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"role": "assistant",
|
| 155 |
+
"content": "请及时就医。",
|
| 156 |
+
"reasoning_content": "现在开始组织回答。",
|
| 157 |
+
"tool_calls": []
|
| 158 |
+
}
|
| 159 |
+
]
|
encoding/tests/test_input_4.json
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"role": "system",
|
| 4 |
+
"content": "该助手为DeepSeek-V3，由深度求索公司创造。\n今天是2025年10月17日，星期五。"
|
| 5 |
+
},
|
| 6 |
+
{
|
| 7 |
+
"role": "latest_reminder",
|
| 8 |
+
"content": "2024-11-15,上海市,App,中文"
|
| 9 |
+
},
|
| 10 |
+
{
|
| 11 |
+
"role": "user",
|
| 12 |
+
"content": "热海大滚锅是世界著名温泉吗"
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"role": "assistant",
|
| 16 |
+
"content": "关于热海大滚锅是否是世界著名温泉，可以这样总结：它在**中国乃至全球的地热奇观中占有重要地位**，但“世界著名”的称号更侧重于它作为**独特的地质现象和旅游景点**，而非普遍的温泉疗养体验。\n\n为了让你快速了解，我整理了一个简要的表格：\n\n| 维度 | 热海大滚锅的地位与特点 |\n| :--- | :--- |\n| **地质奇观** | **世界罕见**的高温地热系统，泉眼多、水温高、形态多样。 |\n| **国内声誉** | **中国三大地热区之一**，**国家5A级旅游景区**，自古闻名（徐霞客曾游历并记载）。 |\n| **国际知名度** | 在一些旅游平台被提及为“世界六大温泉”之一，但此说法流传不广，其国际声誉更多建立在地质独特性上。 |\n| **核心体验** | **观赏地热奇观**（如97℃沸腾的“大滚锅”）、**体验温泉煮鸡蛋**。 |\n\n### 💡 游玩攻略与温馨提示\n\n如果你计划前往热海大滚锅，这里有一些实用信息供你参考：\n\n- **门票与开放时间**：\n - **门票**：景区门票约为**50元/人**。如果选择包含温泉沐浴的套餐，价格会更高，例如约**288元**。\n - **开放时间**：景区一般**08:00-18:00**开放，但具体时间可能变动，建议提前核实。\n\n- **特色体验**：\n - **温泉煮鸡蛋**：这几乎是必试项目。可以在景区门口购买用草绳串起的生鸡蛋（约5-8元/串），然后到“大滚锅”旁的指定区域蒸煮，几分钟便可熟食，趣味十足。\n - **金汤足浴**：可以直接用从“大滚锅”流出的温泉水泡脚，缓解旅途疲劳。\n\n- **注意事项**：\n - **安全第一**：“大滚锅”水温极高，务必遵守游览规则，在指定区域内观赏，切勿随意触碰泉水。\n - **规划行程**：建议为热海景区预留**3-4小时**的游览时间。景区内步道不走回头路，出入口有观光车接送。\n\n希望这些信息能帮助你更好地了解热海大滚锅。如果你对腾冲的其他景点或者行程规划有更多疑问，我很乐意提供进一步的信息。",
|
| 17 |
+
"mask": 1
|
| 18 |
+
},
|
| 19 |
+
{
|
| 20 |
+
"role": "user",
|
| 21 |
+
"content": "世界著名温泉有哪些",
|
| 22 |
+
"task": "action"
|
| 23 |
+
},
|
| 24 |
+
{
|
| 25 |
+
"role": "assistant",
|
| 26 |
+
"content": "Search"
|
| 27 |
+
}
|
| 28 |
+
]
|
encoding/tests/test_output_1.txt
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
<｜begin▁of▁sentence｜>You are a helpful assistant.
|
| 2 |
+
|
| 3 |
+
## Tools
|
| 4 |
+
|
| 5 |
+
You have access to a set of tools to help answer the user's question. You can invoke tools by writing a "<｜DSML｜tool_calls>" block like the following:
|
| 6 |
+
|
| 7 |
+
<｜DSML｜tool_calls>
|
| 8 |
+
<｜DSML｜invoke name="$TOOL_NAME">
|
| 9 |
+
<｜DSML｜parameter name="$PARAMETER_NAME" string="true|false">$PARAMETER_VALUE</｜DSML｜parameter>
|
| 10 |
+
...
|
| 11 |
+
</｜DSML｜invoke>
|
| 12 |
+
<｜DSML｜invoke name="$TOOL_NAME2">
|
| 13 |
+
...
|
| 14 |
+
</｜DSML｜invoke>
|
| 15 |
+
</｜DSML｜tool_calls>
|
| 16 |
+
|
| 17 |
+
String parameters should be specified as is and set `string="true"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string="false"`.
|
| 18 |
+
|
| 19 |
+
If thinking_mode is enabled (triggered by <think>), you MUST output your complete reasoning inside <think>...</think> BEFORE any tool calls or final response.
|
| 20 |
+
|
| 21 |
+
Otherwise, output directly after </think> with tool calls or final response.
|
| 22 |
+
|
| 23 |
+
### Available Tool Schemas
|
| 24 |
+
|
| 25 |
+
{"name": "get_weather", "description": "Get the weather for a specific location", "parameters": {"type": "object", "properties": {"location": {"type": "string", "description": "The city name"}, "unit": {"type": "string", "enum": ["celsius", "fahrenheit"], "description": "Temperature unit"}}, "required": ["location"]}}
|
| 26 |
+
{"name": "search", "description": "Search the web for information", "parameters": {"type": "object", "properties": {"query": {"type": "string", "description": "Search query"}, "num_results": {"type": "integer", "description": "Number of results to return"}}, "required": ["query"]}}
|
| 27 |
+
|
| 28 |
+
You MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.
|
| 29 |
+
<｜User｜>What's the weather in Beijing?<｜Assistant｜><think>The user wants to know the weather in Beijing. I should use the get_weather tool.</think>
|
| 30 |
+
|
| 31 |
+
<｜DSML｜tool_calls>
|
| 32 |
+
<｜DSML｜invoke name="get_weather">
|
| 33 |
+
<｜DSML｜parameter name="location" string="true">Beijing</｜DSML｜parameter>
|
| 34 |
+
<｜DSML｜parameter name="unit" string="true">celsius</｜DSML｜parameter>
|
| 35 |
+
</｜DSML｜invoke>
|
| 36 |
+
</｜DSML｜tool_calls><｜end▁of▁sentence｜><｜User｜><tool_result>{"temperature": 22, "condition": "sunny", "humidity": 45}</tool_result><｜Assistant｜><think>Got the weather data. Let me format a nice response.</think>The weather in Beijing is currently sunny with a temperature of 22°C and 45% humidity.<｜end▁of▁sentence｜>
|
encoding/tests/test_output_2.txt
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
<｜begin▁of▁sentence｜>You are a helpful assistant.<｜User｜>Hello<｜Assistant｜></think>Hi there! How can I help you?<｜end▁of▁sentence｜><｜User｜>What is the capital of France?<｜Assistant｜><think>The user asks about the capital of France. It is Paris.</think>The capital of France is Paris.<｜end▁of▁sentence｜>
|
encoding/tests/test_output_3.txt
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
<｜begin▁of▁sentence｜>该助手为DeepSeek，由深度求索公司创造。<｜latest_reminder｜>2026-02-21,星期六,广州,App,中文<｜User｜>小柴胡冲剂和布洛芬能一起吃吗？
|
| 2 |
+
|
| 3 |
+
CITATION FORMAT: 【{cursor_id}†L{start_line_id}(-L{end_line_id})?】
|
| 4 |
+
|
| 5 |
+
## Tools
|
| 6 |
+
|
| 7 |
+
You have access to a set of tools to help answer the user's question. You can invoke tools by writing a "<｜DSML｜tool_calls>" block like the following:
|
| 8 |
+
|
| 9 |
+
<｜DSML｜tool_calls>
|
| 10 |
+
<｜DSML｜invoke name="$TOOL_NAME">
|
| 11 |
+
<｜DSML｜parameter name="$PARAMETER_NAME" string="true|false">$PARAMETER_VALUE</｜DSML｜parameter>
|
| 12 |
+
...
|
| 13 |
+
</｜DSML｜invoke>
|
| 14 |
+
<｜DSML｜invoke name="$TOOL_NAME2">
|
| 15 |
+
...
|
| 16 |
+
</｜DSML｜invoke>
|
| 17 |
+
</｜DSML｜tool_calls>
|
| 18 |
+
|
| 19 |
+
String parameters should be specified as is and set `string="true"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string="false"`.
|
| 20 |
+
|
| 21 |
+
If thinking_mode is enabled (triggered by <think>), you MUST output your complete reasoning inside <think>...</think> BEFORE any tool calls or final response.
|
| 22 |
+
|
| 23 |
+
Otherwise, output directly after </think> with tool calls or final response.
|
| 24 |
+
|
| 25 |
+
### Available Tool Schemas
|
| 26 |
+
|
| 27 |
+
{"name": "search", "description": "Web search. Split multiple queries with '||'.", "parameters": {"type": "object", "properties": {"queries": {"type": "string", "description": "query1||query2"}}, "required": ["queries"], "additionalProperties": false, "$schema": "http://json-schema.org/draft-07/schema#"}}
|
| 28 |
+
{"name": "open", "description": "Batch open IDs (format 【{id}†...】) or URLs.", "parameters": {"type": "object", "properties": {"open_list": {"type": "array", "items": {"type": "object", "properties": {"id": {"description": "ID or URL", "anyOf": [{"type": "integer"}, {"type": "string"}], "default": -1}, "cursor": {"type": "integer", "description": "", "default": -1}, "loc": {"type": "integer", "description": "Start line", "default": -1}, "num_lines": {"type": "integer", "description": "", "default": -1}, "view_source": {"type": "boolean", "description": "", "default": false}}, "additionalProperties": false}, "description": ""}}, "required": ["open_list"], "additionalProperties": false, "$schema": "http://json-schema.org/draft-07/schema#"}}
|
| 29 |
+
{"name": "find", "description": "Find exact text pattern in pages.", "parameters": {"type": "object", "properties": {"find_list": {"type": "array", "items": {"type": "object", "properties": {"pattern": {"type": "string", "description": ""}, "cursor": {"type": "integer", "description": "", "default": -1}}, "required": ["pattern"], "additionalProperties": false}, "description": ""}}, "required": ["find_list"], "additionalProperties": false, "$schema": "http://json-schema.org/draft-07/schema#"}}
|
| 30 |
+
|
| 31 |
+
You MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.
|
| 32 |
+
<｜Assistant｜><think>用户想知道小柴胡冲剂和布洛芬能否一起服用。</think>
|
| 33 |
+
|
| 34 |
+
<｜DSML｜tool_calls>
|
| 35 |
+
<｜DSML｜invoke name="search">
|
| 36 |
+
<｜DSML｜parameter name="queries" string="true">小柴胡冲剂 布洛芬 相互作用 一起吃</｜DSML｜parameter>
|
| 37 |
+
</｜DSML｜invoke>
|
| 38 |
+
</｜DSML｜tool_calls><｜end▁of▁sentence｜><｜User｜><tool_result>[0]</tool_result><｜Assistant｜><think>现在开始组织回答。</think>请及时就医。<｜end▁of▁sentence｜>
|
encoding/tests/test_output_4.txt
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
<｜begin▁of▁sentence｜>该助手为DeepSeek-V3，由深度求索公司创造。
|
| 2 |
+
今天是2025年10月17日，星期五。<｜latest_reminder｜>2024-11-15,上海市,App,中文<｜User｜>热海大滚锅是世界著名温泉吗<｜Assistant｜></think>关于热海大滚锅是否是世界著名温泉，可以这样总结：它在**中国乃至全球的地热奇观中占有重要地位**，但“世界著名”的称号更侧重于它作为**独特的地质现象和旅游景点**，而非普遍的温泉疗养体验。
|
| 3 |
+
|
| 4 |
+
为了让你快速了解，我整理了一个简要的表格：
|
| 5 |
+
|
| 6 |
+
| 维度 | 热海大滚锅的地位与特点 |
|
| 7 |
+
| :--- | :--- |
|
| 8 |
+
| **地质奇观** | **世界罕见**的高温地热系统，泉眼多、水温高、形态多样。 |
|
| 9 |
+
| **国内声誉** | **中国三大地热区之一**，**国家5A级旅游景区**，自古闻名（徐霞客曾游历并记载）。 |
|
| 10 |
+
| **国际知名度** | 在一些旅游平台被提及为“世界六大温泉”之一，但此说法流传不广，其国际声誉更多建立在地质独特性上。 |
|
| 11 |
+
| **核心体验** | **观赏地热奇观**（如97℃沸腾的“大滚锅”）、**体验温泉煮鸡蛋**。 |
|
| 12 |
+
|
| 13 |
+
### 💡 游玩攻略与温馨提示
|
| 14 |
+
|
| 15 |
+
如果你计划前往热海大滚锅，这里有一些实用信息供你参考：
|
| 16 |
+
|
| 17 |
+
- **门票与开放时间**：
|
| 18 |
+
- **门票**：景区门票约为**50元/人**。如果选择包含温泉沐浴的套餐，价格会更高，例如约**288元**。
|
| 19 |
+
- **开放时间**：景区一般**08:00-18:00**开放，但具体时间可能变动，建议提前核实。
|
| 20 |
+
|
| 21 |
+
- **特色体验**：
|
| 22 |
+
- **温泉煮鸡蛋**：这几乎是必试项目。可以在景区门口购买用草绳串起的生鸡蛋（约5-8元/串），然后到“大滚锅”旁的指定区域蒸煮，几分钟便可熟食，趣味十足。
|
| 23 |
+
- **金汤足浴**：可以直接用从“大滚锅”流出的温泉水泡脚，缓解旅途疲劳。
|
| 24 |
+
|
| 25 |
+
- **注意事项**：
|
| 26 |
+
- **安全第一**：“大滚锅”水温极高，务必遵守游览规则，在指定区域内观赏，切勿随意触碰泉水。
|
| 27 |
+
- **规划行程**：建议为热海景区预留**3-4小时**的游览时间。景区内步道不走回头路，出入口有观光车接送。
|
| 28 |
+
|
| 29 |
+
希望这些信息能帮助你更好地了解热海大滚锅。如果你对腾冲的其他景点或者行程规划有更多疑问，我很乐意提供进一步的信息。<｜end▁of▁sentence｜><｜User｜>世界著名温泉有哪些<｜Assistant｜></think><｜action｜>Search<｜end▁of▁sentence｜>
|
generation_config.json
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_from_model_config": true,
|
| 3 |
+
"bos_token_id": 0,
|
| 4 |
+
"eos_token_id": [
|
| 5 |
+
1,
|
| 6 |
+
128803
|
| 7 |
+
],
|
| 8 |
+
"do_sample": true,
|
| 9 |
+
"temperature": 1.0,
|
| 10 |
+
"top_p": 1.0,
|
| 11 |
+
"transformers_version": "4.46.3"
|
| 12 |
+
}
|
jang_config.json
ADDED
|
@@ -0,0 +1,386 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"weight_format": "mxtq",
|
| 3 |
+
"profile": "JANGTQ-Q2.3",
|
| 4 |
+
"variant": "fractional-power2",
|
| 5 |
+
"mxtq_seed": 42,
|
| 6 |
+
"drop_mtp": true,
|
| 7 |
+
"critical_f32_preserved": true,
|
| 8 |
+
"dsv4_runtime_requirements": {
|
| 9 |
+
"limited_swiglu_tq_patch": true,
|
| 10 |
+
"generic_mlx_sinks": false,
|
| 11 |
+
"native_cache_schema": "deepseek_v4_v7",
|
| 12 |
+
"generic_turboquant_kv": false
|
| 13 |
+
},
|
| 14 |
+
"quantization": {
|
| 15 |
+
"method": "affine+mxtq",
|
| 16 |
+
"routed_experts": {
|
| 17 |
+
"codec": "mxtq",
|
| 18 |
+
"target": "q2.3",
|
| 19 |
+
"effective_bits": 2.2946,
|
| 20 |
+
"bit_plan": {
|
| 21 |
+
"codec": "mxtq",
|
| 22 |
+
"target": "q2.3",
|
| 23 |
+
"effective_bits": 2.2946,
|
| 24 |
+
"bits_used": [
|
| 25 |
+
2,
|
| 26 |
+
4
|
| 27 |
+
],
|
| 28 |
+
"group_count": 129,
|
| 29 |
+
"bit_counts": {
|
| 30 |
+
"2": 110,
|
| 31 |
+
"4": 19
|
| 32 |
+
},
|
| 33 |
+
"assignment": {
|
| 34 |
+
"layer_0.down_proj": 4,
|
| 35 |
+
"layer_0.gate_proj": 4,
|
| 36 |
+
"layer_0.up_proj": 4,
|
| 37 |
+
"layer_1.down_proj": 4,
|
| 38 |
+
"layer_1.gate_proj": 4,
|
| 39 |
+
"layer_1.up_proj": 4,
|
| 40 |
+
"layer_2.down_proj": 4,
|
| 41 |
+
"layer_2.gate_proj": 4,
|
| 42 |
+
"layer_2.up_proj": 4,
|
| 43 |
+
"layer_3.down_proj": 4,
|
| 44 |
+
"layer_3.gate_proj": 2,
|
| 45 |
+
"layer_3.up_proj": 2,
|
| 46 |
+
"layer_4.down_proj": 4,
|
| 47 |
+
"layer_4.gate_proj": 2,
|
| 48 |
+
"layer_4.up_proj": 2,
|
| 49 |
+
"layer_5.down_proj": 4,
|
| 50 |
+
"layer_5.gate_proj": 2,
|
| 51 |
+
"layer_5.up_proj": 2,
|
| 52 |
+
"layer_6.down_proj": 4,
|
| 53 |
+
"layer_6.gate_proj": 2,
|
| 54 |
+
"layer_6.up_proj": 2,
|
| 55 |
+
"layer_7.down_proj": 4,
|
| 56 |
+
"layer_7.gate_proj": 2,
|
| 57 |
+
"layer_7.up_proj": 2,
|
| 58 |
+
"layer_8.down_proj": 4,
|
| 59 |
+
"layer_8.gate_proj": 2,
|
| 60 |
+
"layer_8.up_proj": 2,
|
| 61 |
+
"layer_9.down_proj": 4,
|
| 62 |
+
"layer_9.gate_proj": 2,
|
| 63 |
+
"layer_9.up_proj": 2,
|
| 64 |
+
"layer_10.down_proj": 4,
|
| 65 |
+
"layer_10.gate_proj": 2,
|
| 66 |
+
"layer_10.up_proj": 2,
|
| 67 |
+
"layer_11.down_proj": 4,
|
| 68 |
+
"layer_11.gate_proj": 2,
|
| 69 |
+
"layer_11.up_proj": 2,
|
| 70 |
+
"layer_12.down_proj": 4,
|
| 71 |
+
"layer_12.gate_proj": 2,
|
| 72 |
+
"layer_12.up_proj": 2,
|
| 73 |
+
"layer_13.down_proj": 2,
|
| 74 |
+
"layer_13.gate_proj": 2,
|
| 75 |
+
"layer_13.up_proj": 2,
|
| 76 |
+
"layer_14.down_proj": 2,
|
| 77 |
+
"layer_14.gate_proj": 2,
|
| 78 |
+
"layer_14.up_proj": 2,
|
| 79 |
+
"layer_15.down_proj": 2,
|
| 80 |
+
"layer_15.gate_proj": 2,
|
| 81 |
+
"layer_15.up_proj": 2,
|
| 82 |
+
"layer_16.down_proj": 2,
|
| 83 |
+
"layer_16.gate_proj": 2,
|
| 84 |
+
"layer_16.up_proj": 2,
|
| 85 |
+
"layer_17.down_proj": 2,
|
| 86 |
+
"layer_17.gate_proj": 2,
|
| 87 |
+
"layer_17.up_proj": 2,
|
| 88 |
+
"layer_18.down_proj": 2,
|
| 89 |
+
"layer_18.gate_proj": 2,
|
| 90 |
+
"layer_18.up_proj": 2,
|
| 91 |
+
"layer_19.down_proj": 2,
|
| 92 |
+
"layer_19.gate_proj": 2,
|
| 93 |
+
"layer_19.up_proj": 2,
|
| 94 |
+
"layer_20.down_proj": 2,
|
| 95 |
+
"layer_20.gate_proj": 2,
|
| 96 |
+
"layer_20.up_proj": 2,
|
| 97 |
+
"layer_21.down_proj": 2,
|
| 98 |
+
"layer_21.gate_proj": 2,
|
| 99 |
+
"layer_21.up_proj": 2,
|
| 100 |
+
"layer_22.down_proj": 2,
|
| 101 |
+
"layer_22.gate_proj": 2,
|
| 102 |
+
"layer_22.up_proj": 2,
|
| 103 |
+
"layer_23.down_proj": 2,
|
| 104 |
+
"layer_23.gate_proj": 2,
|
| 105 |
+
"layer_23.up_proj": 2,
|
| 106 |
+
"layer_24.down_proj": 2,
|
| 107 |
+
"layer_24.gate_proj": 2,
|
| 108 |
+
"layer_24.up_proj": 2,
|
| 109 |
+
"layer_25.down_proj": 2,
|
| 110 |
+
"layer_25.gate_proj": 2,
|
| 111 |
+
"layer_25.up_proj": 2,
|
| 112 |
+
"layer_26.down_proj": 2,
|
| 113 |
+
"layer_26.gate_proj": 2,
|
| 114 |
+
"layer_26.up_proj": 2,
|
| 115 |
+
"layer_27.down_proj": 2,
|
| 116 |
+
"layer_27.gate_proj": 2,
|
| 117 |
+
"layer_27.up_proj": 2,
|
| 118 |
+
"layer_28.down_proj": 2,
|
| 119 |
+
"layer_28.gate_proj": 2,
|
| 120 |
+
"layer_28.up_proj": 2,
|
| 121 |
+
"layer_29.down_proj": 2,
|
| 122 |
+
"layer_29.gate_proj": 2,
|
| 123 |
+
"layer_29.up_proj": 2,
|
| 124 |
+
"layer_30.down_proj": 2,
|
| 125 |
+
"layer_30.gate_proj": 2,
|
| 126 |
+
"layer_30.up_proj": 2,
|
| 127 |
+
"layer_31.down_proj": 2,
|
| 128 |
+
"layer_31.gate_proj": 2,
|
| 129 |
+
"layer_31.up_proj": 2,
|
| 130 |
+
"layer_32.down_proj": 2,
|
| 131 |
+
"layer_32.gate_proj": 2,
|
| 132 |
+
"layer_32.up_proj": 2,
|
| 133 |
+
"layer_33.down_proj": 2,
|
| 134 |
+
"layer_33.gate_proj": 2,
|
| 135 |
+
"layer_33.up_proj": 2,
|
| 136 |
+
"layer_34.down_proj": 2,
|
| 137 |
+
"layer_34.gate_proj": 2,
|
| 138 |
+
"layer_34.up_proj": 2,
|
| 139 |
+
"layer_35.down_proj": 2,
|
| 140 |
+
"layer_35.gate_proj": 2,
|
| 141 |
+
"layer_35.up_proj": 2,
|
| 142 |
+
"layer_36.down_proj": 2,
|
| 143 |
+
"layer_36.gate_proj": 2,
|
| 144 |
+
"layer_36.up_proj": 2,
|
| 145 |
+
"layer_37.down_proj": 2,
|
| 146 |
+
"layer_37.gate_proj": 2,
|
| 147 |
+
"layer_37.up_proj": 2,
|
| 148 |
+
"layer_38.down_proj": 2,
|
| 149 |
+
"layer_38.gate_proj": 2,
|
| 150 |
+
"layer_38.up_proj": 2,
|
| 151 |
+
"layer_39.down_proj": 2,
|
| 152 |
+
"layer_39.gate_proj": 2,
|
| 153 |
+
"layer_39.up_proj": 2,
|
| 154 |
+
"layer_40.down_proj": 2,
|
| 155 |
+
"layer_40.gate_proj": 2,
|
| 156 |
+
"layer_40.up_proj": 2,
|
| 157 |
+
"layer_41.down_proj": 2,
|
| 158 |
+
"layer_41.gate_proj": 2,
|
| 159 |
+
"layer_41.up_proj": 2,
|
| 160 |
+
"layer_42.down_proj": 2,
|
| 161 |
+
"layer_42.gate_proj": 2,
|
| 162 |
+
"layer_42.up_proj": 2
|
| 163 |
+
},
|
| 164 |
+
"assignment_unit": "layer.projection, shared across all 256 routed experts"
|
| 165 |
+
}
|
| 166 |
+
},
|
| 167 |
+
"non_routed": {
|
| 168 |
+
"bits": 8,
|
| 169 |
+
"codec": "affine",
|
| 170 |
+
"group_size": 32
|
| 171 |
+
},
|
| 172 |
+
"critical_control_tensors": "source-f32"
|
| 173 |
+
},
|
| 174 |
+
"source_model": "DSV4F-FP",
|
| 175 |
+
"source_config": {
|
| 176 |
+
"n_routed_experts": 256,
|
| 177 |
+
"num_hidden_layers": 43,
|
| 178 |
+
"n_hash_layers": 3
|
| 179 |
+
},
|
| 180 |
+
"routed_expert_bit_plan": {
|
| 181 |
+
"codec": "mxtq",
|
| 182 |
+
"target": "q2.3",
|
| 183 |
+
"effective_bits": 2.2946,
|
| 184 |
+
"bits_used": [
|
| 185 |
+
2,
|
| 186 |
+
4
|
| 187 |
+
],
|
| 188 |
+
"group_count": 129,
|
| 189 |
+
"bit_counts": {
|
| 190 |
+
"2": 110,
|
| 191 |
+
"4": 19
|
| 192 |
+
},
|
| 193 |
+
"assignment": {
|
| 194 |
+
"layer_0.down_proj": 4,
|
| 195 |
+
"layer_0.gate_proj": 4,
|
| 196 |
+
"layer_0.up_proj": 4,
|
| 197 |
+
"layer_1.down_proj": 4,
|
| 198 |
+
"layer_1.gate_proj": 4,
|
| 199 |
+
"layer_1.up_proj": 4,
|
| 200 |
+
"layer_2.down_proj": 4,
|
| 201 |
+
"layer_2.gate_proj": 4,
|
| 202 |
+
"layer_2.up_proj": 4,
|
| 203 |
+
"layer_3.down_proj": 4,
|
| 204 |
+
"layer_3.gate_proj": 2,
|
| 205 |
+
"layer_3.up_proj": 2,
|
| 206 |
+
"layer_4.down_proj": 4,
|
| 207 |
+
"layer_4.gate_proj": 2,
|
| 208 |
+
"layer_4.up_proj": 2,
|
| 209 |
+
"layer_5.down_proj": 4,
|
| 210 |
+
"layer_5.gate_proj": 2,
|
| 211 |
+
"layer_5.up_proj": 2,
|
| 212 |
+
"layer_6.down_proj": 4,
|
| 213 |
+
"layer_6.gate_proj": 2,
|
| 214 |
+
"layer_6.up_proj": 2,
|
| 215 |
+
"layer_7.down_proj": 4,
|
| 216 |
+
"layer_7.gate_proj": 2,
|
| 217 |
+
"layer_7.up_proj": 2,
|
| 218 |
+
"layer_8.down_proj": 4,
|
| 219 |
+
"layer_8.gate_proj": 2,
|
| 220 |
+
"layer_8.up_proj": 2,
|
| 221 |
+
"layer_9.down_proj": 4,
|
| 222 |
+
"layer_9.gate_proj": 2,
|
| 223 |
+
"layer_9.up_proj": 2,
|
| 224 |
+
"layer_10.down_proj": 4,
|
| 225 |
+
"layer_10.gate_proj": 2,
|
| 226 |
+
"layer_10.up_proj": 2,
|
| 227 |
+
"layer_11.down_proj": 4,
|
| 228 |
+
"layer_11.gate_proj": 2,
|
| 229 |
+
"layer_11.up_proj": 2,
|
| 230 |
+
"layer_12.down_proj": 4,
|
| 231 |
+
"layer_12.gate_proj": 2,
|
| 232 |
+
"layer_12.up_proj": 2,
|
| 233 |
+
"layer_13.down_proj": 2,
|
| 234 |
+
"layer_13.gate_proj": 2,
|
| 235 |
+
"layer_13.up_proj": 2,
|
| 236 |
+
"layer_14.down_proj": 2,
|
| 237 |
+
"layer_14.gate_proj": 2,
|
| 238 |
+
"layer_14.up_proj": 2,
|
| 239 |
+
"layer_15.down_proj": 2,
|
| 240 |
+
"layer_15.gate_proj": 2,
|
| 241 |
+
"layer_15.up_proj": 2,
|
| 242 |
+
"layer_16.down_proj": 2,
|
| 243 |
+
"layer_16.gate_proj": 2,
|
| 244 |
+
"layer_16.up_proj": 2,
|
| 245 |
+
"layer_17.down_proj": 2,
|
| 246 |
+
"layer_17.gate_proj": 2,
|
| 247 |
+
"layer_17.up_proj": 2,
|
| 248 |
+
"layer_18.down_proj": 2,
|
| 249 |
+
"layer_18.gate_proj": 2,
|
| 250 |
+
"layer_18.up_proj": 2,
|
| 251 |
+
"layer_19.down_proj": 2,
|
| 252 |
+
"layer_19.gate_proj": 2,
|
| 253 |
+
"layer_19.up_proj": 2,
|
| 254 |
+
"layer_20.down_proj": 2,
|
| 255 |
+
"layer_20.gate_proj": 2,
|
| 256 |
+
"layer_20.up_proj": 2,
|
| 257 |
+
"layer_21.down_proj": 2,
|
| 258 |
+
"layer_21.gate_proj": 2,
|
| 259 |
+
"layer_21.up_proj": 2,
|
| 260 |
+
"layer_22.down_proj": 2,
|
| 261 |
+
"layer_22.gate_proj": 2,
|
| 262 |
+
"layer_22.up_proj": 2,
|
| 263 |
+
"layer_23.down_proj": 2,
|
| 264 |
+
"layer_23.gate_proj": 2,
|
| 265 |
+
"layer_23.up_proj": 2,
|
| 266 |
+
"layer_24.down_proj": 2,
|
| 267 |
+
"layer_24.gate_proj": 2,
|
| 268 |
+
"layer_24.up_proj": 2,
|
| 269 |
+
"layer_25.down_proj": 2,
|
| 270 |
+
"layer_25.gate_proj": 2,
|
| 271 |
+
"layer_25.up_proj": 2,
|
| 272 |
+
"layer_26.down_proj": 2,
|
| 273 |
+
"layer_26.gate_proj": 2,
|
| 274 |
+
"layer_26.up_proj": 2,
|
| 275 |
+
"layer_27.down_proj": 2,
|
| 276 |
+
"layer_27.gate_proj": 2,
|
| 277 |
+
"layer_27.up_proj": 2,
|
| 278 |
+
"layer_28.down_proj": 2,
|
| 279 |
+
"layer_28.gate_proj": 2,
|
| 280 |
+
"layer_28.up_proj": 2,
|
| 281 |
+
"layer_29.down_proj": 2,
|
| 282 |
+
"layer_29.gate_proj": 2,
|
| 283 |
+
"layer_29.up_proj": 2,
|
| 284 |
+
"layer_30.down_proj": 2,
|
| 285 |
+
"layer_30.gate_proj": 2,
|
| 286 |
+
"layer_30.up_proj": 2,
|
| 287 |
+
"layer_31.down_proj": 2,
|
| 288 |
+
"layer_31.gate_proj": 2,
|
| 289 |
+
"layer_31.up_proj": 2,
|
| 290 |
+
"layer_32.down_proj": 2,
|
| 291 |
+
"layer_32.gate_proj": 2,
|
| 292 |
+
"layer_32.up_proj": 2,
|
| 293 |
+
"layer_33.down_proj": 2,
|
| 294 |
+
"layer_33.gate_proj": 2,
|
| 295 |
+
"layer_33.up_proj": 2,
|
| 296 |
+
"layer_34.down_proj": 2,
|
| 297 |
+
"layer_34.gate_proj": 2,
|
| 298 |
+
"layer_34.up_proj": 2,
|
| 299 |
+
"layer_35.down_proj": 2,
|
| 300 |
+
"layer_35.gate_proj": 2,
|
| 301 |
+
"layer_35.up_proj": 2,
|
| 302 |
+
"layer_36.down_proj": 2,
|
| 303 |
+
"layer_36.gate_proj": 2,
|
| 304 |
+
"layer_36.up_proj": 2,
|
| 305 |
+
"layer_37.down_proj": 2,
|
| 306 |
+
"layer_37.gate_proj": 2,
|
| 307 |
+
"layer_37.up_proj": 2,
|
| 308 |
+
"layer_38.down_proj": 2,
|
| 309 |
+
"layer_38.gate_proj": 2,
|
| 310 |
+
"layer_38.up_proj": 2,
|
| 311 |
+
"layer_39.down_proj": 2,
|
| 312 |
+
"layer_39.gate_proj": 2,
|
| 313 |
+
"layer_39.up_proj": 2,
|
| 314 |
+
"layer_40.down_proj": 2,
|
| 315 |
+
"layer_40.gate_proj": 2,
|
| 316 |
+
"layer_40.up_proj": 2,
|
| 317 |
+
"layer_41.down_proj": 2,
|
| 318 |
+
"layer_41.gate_proj": 2,
|
| 319 |
+
"layer_41.up_proj": 2,
|
| 320 |
+
"layer_42.down_proj": 2,
|
| 321 |
+
"layer_42.gate_proj": 2,
|
| 322 |
+
"layer_42.up_proj": 2
|
| 323 |
+
},
|
| 324 |
+
"assignment_unit": "layer.projection, shared across all 256 routed experts"
|
| 325 |
+
},
|
| 326 |
+
"mxtq_bits": {
|
| 327 |
+
"routed_expert": 2,
|
| 328 |
+
"attention": 8,
|
| 329 |
+
"shared_expert": 8,
|
| 330 |
+
"compressor": 8,
|
| 331 |
+
"indexer": 8,
|
| 332 |
+
"embed_tokens": 8,
|
| 333 |
+
"lm_head": 8,
|
| 334 |
+
"norms_router_hc": 16
|
| 335 |
+
},
|
| 336 |
+
"model_family": "deepseek_v4",
|
| 337 |
+
"chat": {
|
| 338 |
+
"encoder": "encoding_dsv4",
|
| 339 |
+
"encoder_fn": "encode_messages",
|
| 340 |
+
"chat_template_source": "builtin_encoding_module",
|
| 341 |
+
"has_tokenizer_chat_template": false,
|
| 342 |
+
"bos_token": "<\uff5cbegin\u2581of\u2581sentence\uff5c>",
|
| 343 |
+
"eos_token": "<\uff5cend\u2581of\u2581sentence\uff5c>",
|
| 344 |
+
"bos_token_id": 0,
|
| 345 |
+
"eos_token_id": 1,
|
| 346 |
+
"role_tokens": {
|
| 347 |
+
"user": "<\uff5cUser\uff5c>",
|
| 348 |
+
"assistant": "<\uff5cAssistant\uff5c>",
|
| 349 |
+
"latest_reminder": "<\uff5clatest_reminder\uff5c>"
|
| 350 |
+
},
|
| 351 |
+
"reasoning": {
|
| 352 |
+
"supported": true,
|
| 353 |
+
"modes": [
|
| 354 |
+
"chat",
|
| 355 |
+
"thinking"
|
| 356 |
+
],
|
| 357 |
+
"default_mode": "chat",
|
| 358 |
+
"thinking_start": "<think>",
|
| 359 |
+
"thinking_end": "</think>",
|
| 360 |
+
"reasoning_effort_levels": [
|
| 361 |
+
"max",
|
| 362 |
+
"high",
|
| 363 |
+
null
|
| 364 |
+
],
|
| 365 |
+
"drop_earlier_reasoning": true
|
| 366 |
+
},
|
| 367 |
+
"tool_calling": {
|
| 368 |
+
"supported": true,
|
| 369 |
+
"parser": "dsml",
|
| 370 |
+
"dsml_token": "\uff5cDSML\uff5c",
|
| 371 |
+
"tool_calls_block": "tool_calls",
|
| 372 |
+
"invoke_block": "invoke",
|
| 373 |
+
"parameter_block": "parameter",
|
| 374 |
+
"tool_output_tag": "tool_result"
|
| 375 |
+
},
|
| 376 |
+
"sampling_defaults": {
|
| 377 |
+
"temperature": 0.6,
|
| 378 |
+
"top_p": 0.95,
|
| 379 |
+
"repetition_penalty": 1.0,
|
| 380 |
+
"repetition_penalty_thinking": 1.0,
|
| 381 |
+
"repetition_penalty_chat": 1.05,
|
| 382 |
+
"max_new_tokens": 4096
|
| 383 |
+
}
|
| 384 |
+
},
|
| 385 |
+
"routed_expert_layout": "prestacked"
|
| 386 |
+
}
|
model-00001-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b9eaa3c8360ea1fed58844ce8b798511f4d6ce29e7cec012aa86deb761861167
|
| 3 |
+
size 1073741952
|
model-00002-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d8f11548fe2ee466c9c0166c5f3c4de892fae14ef65f1909285a09d25f2afeef
|
| 3 |
+
size 1075839337
|
model-00004-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cded486ac428d277d32f987cce7191166c4563520d4f6bea102ecfe3ae984dda
|
| 3 |
+
size 1074790753
|
model-00010-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5118d4b7b6e1edf08952373879ed66ca8f8a6851a654c1b1b32882170b93e7ca
|
| 3 |
+
size 1076888266
|
model-00013-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7befbd59fb34522fbcc919ad500cb351de70ed41c21898e2f9f919ee8efc382d
|
| 3 |
+
size 1076888266
|
model-00015-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bc698cc4bf842c9dd980a60b4f974c271060c2f9d836d36662ee5c69170e62b9
|
| 3 |
+
size 1076888266
|
model-00016-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2b50942d241e028ba8c14535a5b5b6c5b813dea7dc90ba116613683640680e76
|
| 3 |
+
size 1076888266
|
model-00019-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f8f39223f800d884803cbdc0a235a4dfc37ce24bf8a6e2ab2ef7c9438a3e7286
|
| 3 |
+
size 1076888266
|
model-00021-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d502ae6ed1322d7f3d7cbcd5fa55172cd079dee56b8af7211a1a06fdb90331fd
|
| 3 |
+
size 1076888266
|
model-00022-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1e8488f48f9ec8e47e51744a18f8ead512dedda9a490da2894e973f858c0806e
|
| 3 |
+
size 1076888266
|
model-00024-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:87897bc7767bffd794292a5c19a91fa326b33dbb849188ed1cf5d577e7dcc6b8
|
| 3 |
+
size 1075839337
|
model-00027-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:41e4789c93d134526af123bb14fd71dced75eca2797854cb6ebbe0d840fc5ff4
|
| 3 |
+
size 1075839682
|
model-00028-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:45cc733b01b0e1f11548a849842175bf2b11be6346d37a1544112b7d46d20159
|
| 3 |
+
size 1076888266
|
model-00039-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2e03a727abe6e103afd779468620b4f43a6804e209e0eb302536fcb9acf327f5
|
| 3 |
+
size 1075839682
|
model-00040-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3310de5409cfe84745fdf2214bb1b9d50276952dbb6c722b94b65c0b3d1ad445
|
| 3 |
+
size 1076888266
|
model-00043-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:73ee422a34104b1be8a964342b2e40e2fc65d75bf14e64f499c0f642331d7f98
|
| 3 |
+
size 1076888258
|
model-00045-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3ec8da8ae0f36785dc38f07ba4b33ec5492b6882a7b563f322db25b0b649530a
|
| 3 |
+
size 1076888266
|
model-00051-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6fede654ed33e1e808b3a976b60c89e4412fd0f9aecf77ed6bd0da8427eeaa31
|
| 3 |
+
size 1076888266
|
model-00052-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:90817b58a75b0424806f7219cec5d254d8c3b68f65a1773d63f6b3427baa0627
|
| 3 |
+
size 1076888266
|
model-00057-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3335bf013ef8a1a0ba4dd6f03e51001a5ea9076672cc8351992fe144dd13e29b
|
| 3 |
+
size 1076888266
|
model-00058-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bb469caefc65fffc30d5320db8cd2fa8feefabd820e7a64c67fa00604804e753
|
| 3 |
+
size 1074790753
|
model-00060-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:34349707d9be52967a4c18e76bfd00e9929f6bf65772ad35cb0a114a78fce718
|
| 3 |
+
size 1076888258
|
model-00063-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c73cf0f7d8a84bcddbbba644865273fc13ba2422b88053abfb0d33081db3eb14
|
| 3 |
+
size 1076888266
|
model-00065-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:987757c42c8291eb0f7cfc044f1ade0e095aa702206ca60451bc417384a33310
|
| 3 |
+
size 1076888258
|
model-00066-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8ea720fc5b14ef4bf8bf553cf9f3377587f340b420276fe65f9eff4168e999ba
|
| 3 |
+
size 1074790753
|
model-00069-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0ddc73d5bab611a166a3f2b6db212c034f6b892c6045f946a44779ea876ae31c
|
| 3 |
+
size 1076888258
|
model-00071-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:65b568f2c8593d40518e2b2bd12570407a896bd26fc4effc55191291e485e5c2
|
| 3 |
+
size 1076888258
|
model-00074-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:41252c8a0e9ed55ddf50f52a9c71db25225fc8c0b30a417b67a591e45dd20ef6
|
| 3 |
+
size 1016400673
|
model-00082-of-00082.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ad17a57e9a115b4d1be55ea8cf39329dd9311a60d803e30f0a74ca0c3bebb445
|
| 3 |
+
size 33573936
|
model.safetensors.index.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_bos_token": false,
|
| 3 |
+
"add_eos_token": false,
|
| 4 |
+
"bos_token": {
|
| 5 |
+
"__type": "AddedToken",
|
| 6 |
+
"content": "<｜begin▁of▁sentence｜>",
|
| 7 |
+
"lstrip": false,
|
| 8 |
+
"normalized": true,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false
|
| 11 |
+
},
|
| 12 |
+
"clean_up_tokenization_spaces": false,
|
| 13 |
+
"eos_token": {
|
| 14 |
+
"__type": "AddedToken",
|
| 15 |
+
"content": "<｜end▁of▁sentence｜>",
|
| 16 |
+
"lstrip": false,
|
| 17 |
+
"normalized": true,
|
| 18 |
+
"rstrip": false,
|
| 19 |
+
"single_word": false
|
| 20 |
+
},
|
| 21 |
+
"legacy": true,
|
| 22 |
+
"model_max_length": 1048576,
|
| 23 |
+
"pad_token": {
|
| 24 |
+
"__type": "AddedToken",
|
| 25 |
+
"content": "<｜end▁of▁sentence｜>",
|
| 26 |
+
"lstrip": false,
|
| 27 |
+
"normalized": true,
|
| 28 |
+
"rstrip": false,
|
| 29 |
+
"single_word": false
|
| 30 |
+
},
|
| 31 |
+
"sp_model_kwargs": {},
|
| 32 |
+
"unk_token": null,
|
| 33 |
+
"tokenizer_class": "PreTrainedTokenizerFast",
|
| 34 |
+
"eos_token_id": [
|
| 35 |
+
1,
|
| 36 |
+
128803
|
| 37 |
+
]
|
| 38 |
+
}
|