DeepseekR1_V3_tutorial.md

# GPT-4/o1-level Local VSCode Copilot on a Desktop with only 24GB VRAM
# SUMMARY

> **Fed 10, 2025**: Support DeepseekR1 and V3 on single (24GB VRAM)/multi gpu and 382G DRAM, up to 3~64x speedup.<br>

Hi, we're the KTransformers team (formerly known for our local CPU/GPU hybrid inference open source project with DeepSeek-V2).  

We've heard your requests for DeepSeek-R1/V3 support—and we're excited to finally deliver! 
Apologies for the wait, but we've been cooking up something truly amazing!

Today, we're proud to announce that we not only support DeepSeek-R1/V3, as showcased in the video below:  

https://github.com/user-attachments/assets/ebd70bfa-b2c1-4abb-ae3b-296ed38aa285

</p>

- **[NEW!!!] Local 671B DeepSeek-Coder-V3/R1:** Running its Q4_K_M version using only 14GB VRAM and 382GB DRAM.
	- Prefill Speed (tokens/s): 
 		- KTransfermor: 54.21 (32 cores) → 74.362 (dual-socket, 2×32 cores) → 255.26 (optimized AMX-based MoE kernel, V0.3 only) → 286.55 (selectively using 6 experts, V0.3 only)  
 		- Compared to 4.51 tokens/s in llama.cpp with 2×32 cores, achieving up to **63.53× speedup**.  
 	- Decode Speed (tokens/s):  
 		- KTransfermor: 8.73 (32 cores) → 11.26 (dual-socket, 2×32 cores) → 13.69 (selectively using 6 experts, V0.3 only)  
 		- Compared to 4.51 tokens/s in llama.cpp with 2×32 cores, achieving up to **3.03× speedup**.  
  

We also give our upcoming optimizations previews, including an Intel AMX-accelerated kernel and a selective expert activation method, which will significantly enhance performance. With V0.3-preview, we achieve up to 286 tokens/s for prefill, making it up to **64× faster than llama.cpp** for local inference.
The binary distribution is available now and the source code will come ASAP! Check out the wheel package [here](https://github.com/kvcache-ai/ktransformers/releases/download/v0.1.4/ktransformers-0.3.0rc0+cu126torch26fancy-cp311-cp311-linux_x86_64.whl)  


## Prerequisites
We run our best performance tests (V0.2) on <br>
CPU: Intel (R) Xeon (R) Gold 6454S 1T DRAM (2 NUMA nodes) <br>
GPU: 4090D 24G VRAM <br>
## Bench Result
### V0.2
#### Settings
- Model: DeepseekV3-q4km (int4)<br>
- CPU: cpu_model_name: Intel (R) Xeon (R) Gold 6454S, 32 cores per socket, 2 sockets, 2 numa nodes
- GPU: 4090D 24G VRAM
- We test after enough warm up
#### Memory consumption:
  - Single socket: 382G DRAM, at least 14GB VRAM
  - Dual socket: 1T DRAM, at least 14GB VRAM

#### Benchmark Results

"6 experts" case is part of V0.3's preview

| Prompt<br>(500 tokens) | Dual socket Ktrans (6 experts) | Dual socket Ktrans (8 experts) | Single socket Ktrans (6 experts) | Single socket Ktrans (8 experts)| llama.cpp (8 experts) | 
| --- | --- | --- | --- | --- | --- | 
| Prefill token/s | 97.32 | 82.94 | 65.14 | 54.21 | 10.31 |
| Decode token/s | 13.69 | 12.208 | 10.303 | 8.73 |4.51 |

**The highest speedup reaches up to <u>3.03x</u> in decoding and <u>9.44x</u> in prefill.**

### V0.3-Preview
#### Settings
- Model: DeepseekV3-BF16 (online quant into int8 for CPU and int4 for GPU)
- CPU: cpu_model_name: Intel (R) Xeon (R) Gold 6454S, 32 cores per socket, 2 socket, 2 numa nodes
- GPU: (1~4)x 4090D 24GVRAM (requires more VRAM for longer prompt)

#### Memory consumptions:
- 644GB DRAM, at least 14GB VRAM

#### Benchmark results
| Prompt length  | 1K  | 2K  | 4K  | 8K |
|---------------|-----|-----|-----|-----|
| KTrans (8 experts) Prefill token/s |   185.96  |  255.26   |  252.58   |  195.62   |
| KTrans (6 experts) Prefill token/s |   203.70  |  286.55   |  271.08   |  207.20   |

**The prefill of KTrans V0.3 is up to <u>3.45x</u> times faster than KTrans V0.2, and is up to <u>63.53x</u> times faster than llama.cpp.**
**The decoding speed is the same as KTrans V0.2 (6 experts version) so it is omitted**

The main acceleration comes from 
- Intel AMX instruction set and our specially designed cache friendly memory layout
- Expert selection strategy that selects fewer experts based on offline profile results of out of domain data


*From our research on DeepSeekV2, DeepSeekV3 and DeepSeekR1, 
when we slightly decrease the activation experts num in inference, 
the output quality doesn't change. But the speed of decoding and prefill 
is speed up which is inspiring. So our showcase makes use of this finding*

## How to Run
### V0.2 Showcase
#### Single socket version (32 cores)
Our local_chat test command is:
``` shell
git clone https://github.com/kvcache-ai/ktransformers.git
cd ktransformers
numactl -N 1 -m 1 python ./ktransformers/local_chat.py --model_path <your model path> --gguf_path <your gguf path>  --prompt_file <your prompt txt file>  --cpu_infer 33  --cache_lens 1536 
<when you see chat, then press enter to load the text prompt_file>
```
\<your model path\> can be local or set from online hugging face like deepseek-ai/DeepSeek-V3. If online encounters connection problem, try use mirror (hf-mirror.com) <br>
\<your gguf path\> can also be online, but as its large we recommend you download it and quantize the model to what you want <br>
The command numactl -N 1 -m 1 aims to advoid data transfer between numa nodes
#### Dual socket version (64 cores)
Make suer before you install (use install.sh or `make dev_install`), setting the env var `USE_NUMA=1` by `export USE_NUMA=1` (if already installed, reinstall it with this env var set) <br>
Our local_chat test command is:
``` shell
git clone https://github.com/kvcache-ai/ktransformers.git
cd ktransformers
export USE_NUMA=1
make dev_install # or sh ./install.sh
python ./ktransformers/local_chat.py --model_path <your model path> --gguf_path <your gguf path>  --prompt_file <your prompt txt file>  --cpu_infer 65  --cache_lens 1536 
<when you see chat, then press enter to load the text prompt_file>
```
The parameters' meaning is the same. But As we  use dual socket, we set cpu_infer to 65

### V0.3 Showcase
#### Dual socket version (64 cores)
Our local_chat test command is:
``` shell
wget https://github.com/kvcache-ai/ktransformers/releases/download/v0.1.4/ktransformers-0.3.0rc0+cu126torch26fancy-cp311-cp311-linux_x86_64.whl
pip install ./ktransformers-0.3.0rc0+cu126torch26fancy-cp311-cp311-linux_x86_64.whl
python -m ktransformers.local_chat --model_path <your model path> --gguf_path <your gguf path>  --prompt_file <your prompt txt file>  --cpu_infer 65  --cache_lens 1536 
<when you see chat, then press enter to load the text prompt_file>
```
The parameters' meaning is the same with V0.2. But As we  use dual socket, we set cpu_infer to 65

## Some Explanations
1. Also we want to make further use of our two NUMA nodes on Xeon Gold cpu. 
To avoid the cost of data transfer between nodes, we "copy" the critical matrix on 
both nodes which takes more memory consumption but accelerates the prefill and decoding process.
But this method takes huge memory and slow when loading weights, So be patient when loading
and monitor the memory usage. We are going to optimize this huge memory overhead. Stay tuned~ <br>
2. The command args `--cpu_infer 65` specifies how many cores to use (it's ok that it exceeds the physical number, 
but it's not the more the better. Adjust it slightly lower to your actual number of cores)<br>

3. Why CPU/GPU Hybrid Inference?
DeepSeek's MLA operators are highly computationally intensive. While running everything on CPU is possible, offloading the heavy computations to the GPU results in a massive performance boost.  

4. Where Does the Speedup Come From?

   - Expert Offload: Unlike traditional layer-based or KVCache offloading (as seen in llama.cpp), we offload the expert computation to the CPU and MLA/KVCache to GPU, aligning perfectly with DeepSeek’s architecture for optimal efficiency.  
   - Intel AMX Optimization – Our AMX-accelerated kernel is meticulously tuned, running several times faster than existing llama.cpp implementations. We plan to open-source this kernel after cleansing and are considering upstream contributions to llama.cpp.  

5. Why Intel CPUs?
Intel is currently the only CPU vendor that supports AMX-like instructions, which delivers significantly better performance compared to AVX-only alternatives.