Skip to content

GitLab

Commit c009512a authored by Azure-Tang's avatar Azure-Tang
Browse files

Merge branch 'main' into hip

parents c1f13a69 4f22d726
Hide whitespace changes
Inline Side-by-side
Showing with 1036 additions and 59 deletions
doc/basic/note1.md 0 → 100644
View file @ c009512a
# basic-first20
doc/basic/note2.md 0 → 100644
View file @ c009512a
# basic-data_structure
doc/en/DeepseekR1_V3_tutorial.md
View file @ c009512a
<!-- omit in toc --> <!-- omit in toc -->
# GPT-4/o1-level Local VSCode Copilot on a Desktop with only 24GB VRAM # GPT-4/o1-level Local VSCode Copilot on a Desktop with only 24GB VRAM
- [SUMMARY](#summary) - [SUMMARY](#summary)
- [Prerequisites](#prerequisites) - [Show Case Environment](#show-case-environment)
- [Bench Result](#bench-result) - [Bench Result](#bench-result)
- [V0.2](#v02) - [V0.2.1](#v021)
- [Settings](#settings)
- [Memory consumption:](#memory-consumption) - [Memory consumption:](#memory-consumption)
- [Change Log](#change-log)
- [Benchmark Results](#benchmark-results) - [Benchmark Results](#benchmark-results)
- [V0.2](#v02)
- [Settings](#settings)
- [Memory consumption:](#memory-consumption-1)
- [Benchmark Results](#benchmark-results-1)
- [V0.3-Preview](#v03-preview) - [V0.3-Preview](#v03-preview)
- [Settings](#settings-1) - [Settings](#settings-1)
- [Memory consumptions:](#memory-consumptions) - [Memory consumptions:](#memory-consumptions)
- [Benchmark results](#benchmark-results-1) - [Benchmark results](#benchmark-results-2)
- [How to Run](#how-to-run) - [How to Run](#how-to-run)
- [V0.2 Showcase](#v02-showcase) - [v0.2.2 \& v0.2.3 longer context \& FP8 kernel](#v022--v023-longer-context--fp8-kernel)
- [longer context](#longer-context)
- [FP8 kernel](#fp8-kernel)
- [V0.2 \& V0.2.1 Showcase](#v02--v021-showcase)
- [Single socket version (32 cores)](#single-socket-version-32-cores) - [Single socket version (32 cores)](#single-socket-version-32-cores)
- [Dual socket version (64 cores)](#dual-socket-version-64-cores) - [Dual socket version (64 cores)](#dual-socket-version-64-cores)
- [V0.3 Showcase](#v03-showcase) - [V0.3 Showcase](#v03-showcase)
- [Dual socket version (64 cores)](#dual-socket-version-64-cores-1) - [Dual socket version (64 cores)](#dual-socket-version-64-cores-1)
- [Some Explanations](#some-explanations) - [Some Explanations](#some-explanations)
- [Next](#next)
- [Faster](#faster)
- [Easier](#easier)
- [FAQ](#faq) - [FAQ](#faq)
- [R1 No Thinking](#r1-no-thinking) - [R1 No Thinking](#r1-no-thinking)
- [More FAQ](#more-faq) - [More FAQ](#more-faq)
# SUMMARY # SUMMARY
> **Fed 10, 2025**: Support DeepseekR1 and V3 on single (24GB VRAM)/multi gpu and 382G DRAM, up to 3~28x speedup.<br> > **Feb 10, 2025**: Support DeepseekR1 and V3 on single (24GB VRAM)/multi gpu and 382G DRAM, up to 3~28x speedup.<br>
Hi, we're the KTransformers team (formerly known for our local CPU/GPU hybrid inference open source project with DeepSeek-V2). Hi, we're the KTransformers team (formerly known for our local CPU/GPU hybrid inference open source project with DeepSeek-V2).
...@@ -39,23 +49,64 @@ https://github.com/user-attachments/assets/ebd70bfa-b2c1-4abb-ae3b-296ed38aa285 ...@@ -39,23 +49,64 @@ https://github.com/user-attachments/assets/ebd70bfa-b2c1-4abb-ae3b-296ed38aa285
- **[NEW!!!] Local 671B DeepSeek-Coder-V3/R1:** Running its Q4_K_M version using only 14GB VRAM and 382GB DRAM. - **[NEW!!!] Local 671B DeepSeek-Coder-V3/R1:** Running its Q4_K_M version using only 14GB VRAM and 382GB DRAM.
- Prefill Speed (tokens/s): - 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) - KTransformers: 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 10.31 tokens/s in llama.cpp with 2×32 cores, achieving up to **27.79× speedup**. - Compared to 10.31 tokens/s in llama.cpp with 2×32 cores, achieving up to **27.79× speedup**.
- Decode Speed (tokens/s): - 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) - KTransformers: 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**. - 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 **28× faster than llama.cpp** for local inference. 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 **28× 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) 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)
> **Feb 15, 2025**: KTransformers V0.2.1: Longer Context (from 4K to 8K for 24GB VRAM) & Slightly Faster Speed （+15%) (Up to 16 Tokens/s), update docs [here](./doc/en/DeepseekR1_V3_tutorial.md) and [online books](https://kvcache-ai.github.io/ktransformers/).
We speed up the decode and prefill speed a littlt bit. The reason for the limited performance improvement mainly lies in the fact that the inference process is still constrained by the CPU's computational speed and memory bandwidth. The MLA part handled by the GPU accounts for a relatively small proportion.
## Prerequisites Besides the improvements in speed, we've also significantly updated the documentation to enhance usability, including:<br>
- Added Multi-GPU configuration tutorial.
- Consolidated installation guide.
- Add a detailed tutorial on registering extra GPU memory with ExpertMarlin;
## Show Case Environment
We run our best performance tests (V0.2) on <br> We run our best performance tests (V0.2) on <br>
CPU: Intel (R) Xeon (R) Gold 6454S 1T DRAM (2 NUMA nodes) <br> CPU: Intel (R) Xeon (R) Gold 6454S 1T DRAM (2 NUMA nodes) <br>
GPU: 4090D 24G VRAM <br> GPU: 4090D 24G VRAM <br>
Memory: standard DDR5-4800 server DRAM (1 TB) Memory: standard DDR5-4800 server DRAM (1 TB), each socket with 8×DDR5-4800
## Bench Result ## Bench Result
### V0.2.1
- 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: 4090 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
#### Change Log
- Longer Context (from 4K to 8K for 24GB VRAM) and Slightly Faster Speed （+15%):<br>
Integrated the highly efficient Triton MLA Kernel from the fantastic sglang project, enable much longer context length and slightly faster prefill/decode speed
- We suspect that some of the improvements come from the change of hardware platform (4090D->4090)
#### Benchmark Results
"6 experts" case is part of V0.3's preview
| Prompt | hi (2) | 1K (969) | 2K (1930) | 4K (3846) | 8K (7678) |
| --- | --- | --- | --- | --- | --- |
| Output length | 10tokens | 300tokens | 300tokens | 300tokens | 300tokens |
| **6 experts V0.2.0** | | | | | |
| Prefill token/s | 13 | 105 | 102 | 88 | CUDA OOM |
| decode token/s | 16.8 | 15.4 | 14.2 | 13.0 | CUDA OOM |
| **6 experts V0.2.1** | | | | | |
| Prefill token/s | 13 | 111 | 112.5 | 102 **(1.16x speedup)** | 101 |
| decode token/s | 16.8 | 15.9 | 15.4 | 14.9 **(1.15x speedup)** | 13.9 |
| **8 experts V0.2.1** | | | | | |
| Prefill token/s | 12.2 | 88.2 | 88.5 | 81.9 | 80 |
| Decode token/s | 13.4 | 13.5 | 13.4 | 13.2 | 12.4 |
### V0.2 ### V0.2
#### Settings #### Settings
- Model: DeepseekV3-q4km (int4)<br> - Model: DeepseekV3-q4km (int4)<br>
...@@ -106,14 +157,41 @@ the output quality doesn't change. But the speed of decoding and prefill ...@@ -106,14 +157,41 @@ 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* is speed up which is inspiring. So our showcase makes use of this finding*
## How to Run ## How to Run
### V0.2 Showcase ### v0.2.2 & v0.2.3 longer context & FP8 kernel
#### longer context
To use this feature, [install flashinfer](https://github.com/flashinfer-ai/flashinfer) first.
Note: The latest MLA kernel in FlashInfer still has a few minor issues. They are continuously fixing them on the main branch. If you are using FlashInfer, please install it from the main source code.
If you want to use long context(longer than 20K) for prefill, enable the matrix absorption MLA during the prefill phase, which will significantly reduce the size of the kv cache. Modify yaml file like this:
```
- match:
name: "^model\\.layers\\..*\\.self_attn$"
replace:
class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
kwargs:
generate_device: "cuda"
prefill_device: "cuda"
absorb_for_prefill: True # change this to True to enable long context(prefill may slower).
```
If the VRAM is still insufficient, try reducing the `chunk_prefill_size` parameter (default is 8192) to further decrease the intermediate results during chunk prefill.
#### FP8 kernel
The DeepSeek-AI team provides FP8 safetensors for DeepSeek-R1/V3 models. We achieve performance optimization through the following works:
- **FP8 GPU Kernel Integration**: FP8 linear layer acceleration kernels integrated in KTransformers
- **Hybrid Quantization Architecture**:
- Attention and Shared-Expert modules use FP8 precision (enhances computational accuracy)
- Experts modules retain GGML quantization (GGUF format, reside in CPU to save GPU memory)
So those who are persuing the best performance can use the FP8 linear kernel for DeepSeek-V3/R1.
The detailed guide is [here](./fp8_kernel.md).
### V0.2 & V0.2.1 Showcase
#### Single socket version (32 cores) #### Single socket version (32 cores)
Our local_chat test command is: Our local_chat test command is:
``` shell ``` shell
git clone https://github.com/kvcache-ai/ktransformers.git
cd ktransformers
git submodule init
git submodule update
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 --max_new_tokens 1000 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 --max_new_tokens 1000
<when you see chat, then press enter to load the text prompt_file> <when you see chat, then press enter to load the text prompt_file>
``` ```
...@@ -121,24 +199,28 @@ numactl -N 1 -m 1 python ./ktransformers/local_chat.py --model_path <your model ...@@ -121,24 +199,28 @@ numactl -N 1 -m 1 python ./ktransformers/local_chat.py --model_path <your model
`<your gguf path>` can also be online, but as its large we recommend you download it and quantize the model to what you want (notice it's the dir path) <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 (notice it's the dir path) <br>
`--max_new_tokens 1000` is the max output token length. If you find the answer is truncated, you `--max_new_tokens 1000` is the max output token length. If you find the answer is truncated, you
can increase the number for longer answer (But be aware of OOM, and increase it will slow down the generation rate.). can increase the number for longer answer (But be aware of OOM, and increase it will slow down the generation rate.).
<br>
The command numactl -N 1 -m 1 aims to advoid data transfer between numa nodes<br> The command `numactl -N 1 -m 1` aims to advoid data transfer between numa nodes<br>
Attention! If you are testing R1 and it may skip thinking. So you can add arg: `--force_think true`. This is explained in [FAQ](#faq) part Attention! If you are testing R1 and it may skip thinking. So you can add arg: `--force_think true`. This is explained in [FAQ](#faq) part
#### Dual socket version (64 cores) #### 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: Make sure 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). You may check the doc [here](./install.md) for install details. <br>
Test Command:
``` shell ``` shell
git clone https://github.com/kvcache-ai/ktransformers.git # ---For those who have not installed ktransformers---
cd ktransformers # git clone https://github.com/kvcache-ai/ktransformers.git
git submodule init # cd ktransformers
git submodule update # git submodule init
export USE_NUMA=1 # git submodule update
make dev_install # or sh ./install.sh # 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 --max_new_tokens 1000 python ./ktransformers/local_chat.py --model_path <your model path> --gguf_path <your gguf path> --prompt_file <your prompt txt file> --cpu_infer 65 --max_new_tokens 1000
<when you see chat, then press enter to load the text prompt_file> <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 The parameters' meaning is the same. But As we use dual socket, we set cpu_infer to 65
### V0.3 Showcase ### V0.3 Showcase
#### Dual socket version (64 cores) #### Dual socket version (64 cores)
...@@ -170,9 +252,21 @@ DeepSeek's MLA operators are highly computationally intensive. While running eve ...@@ -170,9 +252,21 @@ DeepSeek's MLA operators are highly computationally intensive. While running eve
5. Why Intel CPUs? 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. Intel is currently the only CPU vendor that supports AMX-like instructions, which delivers significantly better performance compared to AVX-only alternatives.
## Next
### Faster
* The FlashInfer (https://github.com/flashinfer-ai/flashinfer) project is releasing an even more efficient fused MLA operator, promising further speedups
* vLLM has explored multi-token prediction in DeepSeek-V3, and support is on our roadmap for even better performance
* We are collaborating with Intel to enhance the AMX kernel (v0.3) and optimize for Xeon6/MRDIMM
### Easier
* Official Docker images to simplify installation
* Fix the server integration for web API access
* Fix the local chat only accepting a single line prompt (currently \n begins generating prompt)
* Support for more quantization types, including the highly requested dynamic quantization from unsloth
Stay tuned for more updates!
## FAQ ## FAQ
### R1 No Thinking ### R1 No Thinking
Attention! If you are testing R1 and it may skip thinking. So you can add arg: `--force_think true`. The detail is in [FAQ](./FAQ.md) part <br> Attention! If you are testing R1 and it may skip thinking. So you can add arg: `--force_think true`. The detail is in [FAQ](./FAQ.md) part <br>
### More FAQ ### More FAQ
[See detail](./FAQ.md) [See detail](./FAQ.md)
\ No newline at end of file
doc/en/Docker.md
View file @ c009512a
...@@ -7,7 +7,7 @@ ...@@ -7,7 +7,7 @@
## Images ## Images
There is a Docker image available for our project, you can pull the docker image by： There is a Docker image available for our project, you can pull the docker image by：
``` ```
docker pull approachingai/ktransformers:0.1.1 docker pull approachingai/ktransformers:0.2.1
``` ```
**Notice**: In this image, we compile the ktransformers in AVX512 instuction CPUs, if your cpu not support AVX512, it is suggested to recompile and install ktransformer in the /workspace/ktransformers directory within the container. **Notice**: In this image, we compile the ktransformers in AVX512 instuction CPUs, if your cpu not support AVX512, it is suggested to recompile and install ktransformer in the /workspace/ktransformers directory within the container.
...@@ -16,14 +16,16 @@ docker pull approachingai/ktransformers:0.1.1 ...@@ -16,14 +16,16 @@ docker pull approachingai/ktransformers:0.1.1
- finish, execute - finish, execute
```bash ```bash
docker build -t approachingai/ktransformers:v0.1.1 . docker build -t approachingai/ktransformers:0.2.1 .
``` ```
## Usage ## Usage
Assuming you have the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) that you can use the GPU in a Docker container. Assuming you have the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) that you can use the GPU in a Docker container.
``` ```
docker run --gpus all -v /path/to/models:/models -p 10002:10002 approachingai/ktransformers:v0.1.1 --port 10002 --gguf_path /models/path/to/gguf_path --model_path /models/path/to/model_path --web True docker run --gpus all -v /path/to/models:/models --name ktransformers -itd approachingai/ktransformers:0.2.1
docker exec -it ktransformers /bin/bash
python -m ktransformers.local_chat --gguf_path /models/path/to/gguf_path --model_path /models/path/to/model_path --cpu_infer 33
``` ```
More operators you can see in the [readme](../../README.md) More operators you can see in the [readme](../../README.md)
\ No newline at end of file
doc/en/FAQ.md
View file @ c009512a
<!-- omit in toc -->
# FAQ # FAQ
- [Install](#install)
- [Q: ImportError: /lib/x86\_64-linux-gnu/libstdc++.so.6: version GLIBCXX\_3.4.32' not found](#q-importerror-libx86_64-linux-gnulibstdcso6-version-glibcxx_3432-not-found)
- [Q: DeepSeek-R1 not outputting initial token](#q-deepseek-r1-not-outputting-initial--token)
- [Usage](#usage)
- [Q: If I got more VRAM than the model's requirement, how can I fully utilize it?](#q-if-i-got-more-vram-than-the-models-requirement-how-can-i-fully-utilize-it)
- [Q: If I don't have enough VRAM, but I have multiple GPUs, how can I utilize them?](#q-if-i-dont-have-enough-vram-but-i-have-multiple-gpus-how-can-i-utilize-them)
- [Q: How to get the best performance?](#q-how-to-get-the-best-performance)
- [Q: My DeepSeek-R1 model is not thinking.](#q-my-deepseek-r1-model-is-not-thinking)
- [Q: Loading gguf error](#q-loading-gguf-error)
- [Q: Version \`GLIBCXX\_3.4.30' not found](#q-version-glibcxx_3430-not-found)
- [Q: When running the bfloat16 moe model, the data shows NaN](#q-when-running-the-bfloat16-moe-model-the-data-shows-nan)
- [Q: Using fp8 prefill very slow.](#q-using-fp8-prefill-very-slow)
- [Q: Possible ways to run graphics cards using volta and turing architectures](#q-possible-ways-to-run-graphics-cards-using-volta-and-turing-architectures)
## Install ## Install
### Q: ImportError: /lib/x86_64-linux-gnu/libstdc++.so.6: version GLIBCXX_3.4.32' not found ### Q: ImportError: /lib/x86_64-linux-gnu/libstdc++.so.6: version GLIBCXX_3.4.32' not found
``` ```
...@@ -25,7 +39,7 @@ from-https://github.com/kvcache-ai/ktransformers/issues/129#issue-2842799552 ...@@ -25,7 +39,7 @@ from-https://github.com/kvcache-ai/ktransformers/issues/129#issue-2842799552
1. local_chat.py: You can increase the context window size by setting `--max_new_tokens` to a larger value. 1. local_chat.py: You can increase the context window size by setting `--max_new_tokens` to a larger value.
2. server: Increase the `--cache_lens' to a larger value. 2. server: Increase the `--cache_lens' to a larger value.
2. Move more weights to the GPU. 2. Move more weights to the GPU.
Refer to the ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-multi-gpu-marlin.yaml Refer to the ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-multi-gpu-4.yaml
```yaml ```yaml
- match: - match:
name: "^model\\.layers\\.([4-10])\\.mlp\\.experts$" # inject experts in layer 4~10 as marlin expert name: "^model\\.layers\\.([4-10])\\.mlp\\.experts$" # inject experts in layer 4~10 as marlin expert
...@@ -39,11 +53,13 @@ from-https://github.com/kvcache-ai/ktransformers/issues/129#issue-2842799552 ...@@ -39,11 +53,13 @@ from-https://github.com/kvcache-ai/ktransformers/issues/129#issue-2842799552
You can modify layer as you want, eg. `name: "^model\\.layers\\.([4-10])\\.mlp\\.experts$"` to `name: "^model\\.layers\\.([4-12])\\.mlp\\.experts$"` to move more weights to the GPU. You can modify layer as you want, eg. `name: "^model\\.layers\\.([4-10])\\.mlp\\.experts$"` to `name: "^model\\.layers\\.([4-12])\\.mlp\\.experts$"` to move more weights to the GPU.
> Note: The first matched rule in yaml will be applied. For example, if you have two rules that match the same layer, only the first rule's replacement will be valid. > Note: The first matched rule in yaml will be applied. For example, if you have two rules that match the same layer, only the first rule's replacement will be valid.
> Note：Currently, executing experts on the GPU will conflict with CUDA Graph. Without CUDA Graph, there will be a significant slowdown. Therefore, unless you have a substantial amount of VRAM (placing a single layer of experts for DeepSeek-V3/R1 on the GPU requires at least 5.6GB of VRAM), we do not recommend enabling this feature. We are actively working on optimization.
> Note KExpertsTorch is untested.
### Q: If I don't have enough VRAM, but I have multiple GPUs, how can I utilize them? ### Q: If I don't have enough VRAM, but I have multiple GPUs, how can I utilize them?
Use the `--optimize_rule_path ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-multi-gpu.yaml` to load the two optimized rule yaml file. You may also use it as an example to write your own 4/8 gpu optimized rule yaml file. Use the `--optimize_config_path ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-multi-gpu.yaml` to load the two optimized rule yaml file. You may also use it as an example to write your own 4/8 gpu optimized rule yaml file.
> Note: The ktransformers' multi-gpu stratigy is pipline, which is not able to speed up the model's inference. It's only for the model's weight distribution. > Note: The ktransformers' multi-gpu stratigy is pipline, which is not able to speed up the model's inference. It's only for the model's weight distribution.
...@@ -53,7 +69,7 @@ You have to set `--cpu_infer` to the number of cores you want to use. The more c ...@@ -53,7 +69,7 @@ You have to set `--cpu_infer` to the number of cores you want to use. The more c
### Q: My DeepSeek-R1 model is not thinking. ### Q: My DeepSeek-R1 model is not thinking.
According to DeepSeek, you need to enforce the model to initiate its response with "\<think>\n" at the beginning of every output by passing the arg `--force_think true `. According to DeepSeek, you need to enforce the model to initiate its response with "\<think>\n" at the beginning of every output by passing the arg `--force_think True `.
### Q: Loading gguf error ### Q: Loading gguf error
...@@ -61,9 +77,91 @@ Make sure you: ...@@ -61,9 +77,91 @@ Make sure you:
1. Have the `gguf` file in the `--gguf_path` directory. 1. Have the `gguf` file in the `--gguf_path` directory.
2. The directory only contains gguf files from one model. If you have multiple models, you need to separate them into different directories. 2. The directory only contains gguf files from one model. If you have multiple models, you need to separate them into different directories.
3. The folder name it self should not end with `.gguf`, eg. `Deep-gguf` is correct, `Deep.gguf` is wrong. 3. The folder name it self should not end with `.gguf`, eg. `Deep-gguf` is correct, `Deep.gguf` is wrong.
4. The file itself is not corrupted; you can verify this by checking that the sha256sum matches the one from huggingface, modelscope, or hf-mirror.
### Q: Version `GLIBCXX_3.4.30' not found ### Q: Version `GLIBCXX_3.4.30' not found
The detailed error: The detailed error:
>ImportError: /mnt/data/miniconda3/envs/xxx/bin/../lib/libstdc++.so.6: version `GLIBCXX_3.4.30' not found (required by /home/xxx/xxx/ktransformers/./cpuinfer_ext.cpython-312-x86_64-linux-gnu.so) >ImportError: /mnt/data/miniconda3/envs/xxx/bin/../lib/libstdc++.so.6: version `GLIBCXX_3.4.30' not found (required by /home/xxx/xxx/ktransformers/./cpuinfer_ext.cpython-312-x86_64-linux-gnu.so)
It may because of your conda env have no this version. Your can first exit your conda env by `conda deactivate` and use `whereis libstdc++.so.6` to find the path. And re enter your conda env and copy the .so by `cp <path of outter libstdc++> <path of your conda env libstdc++>` Running `conda install -c conda-forge libstdcxx-ng` can solve the problem.
### Q: When running the bfloat16 moe model, the data shows NaN
The detailed error:
```shell
Traceback (most recent call last):
File "/root/ktransformers/ktransformers/local_chat.py", line 183, in <module>
fire.Fire(local_chat)
File "/usr/local/lib/python3.10/dist-packages/fire/core.py", line 135, in Fire
component_trace = _Fire(component, args, parsed_flag_args, context, name)
File "/usr/local/lib/python3.10/dist-packages/fire/core.py", line 468, in _Fire
component, remaining_args = _CallAndUpdateTrace(
File "/usr/local/lib/python3.10/dist-packages/fire/core.py", line 684, in _CallAndUpdateTrace
component = fn(*varargs, **kwargs)
File "/root/ktransformers/ktransformers/local_chat.py", line 177, in local_chat
generated = prefill_and_generate(
File "/root/ktransformers/ktransformers/util/utils.py", line 204, in prefill_and_generate
next_token = decode_one_tokens(cuda_graph_runner, next_token.unsqueeze(0), position_ids, cache_position, past_key_values, use_cuda_graph).to(torch_device)
File "/root/ktransformers/ktransformers/util/utils.py", line 128, in decode_one_tokens
next_token = torch.multinomial(probs, num_samples=1).squeeze(1)
RuntimeError: probability tensor contains either `inf`, `nan` or element < 0
```
**SOLUTION**: The issue of running ktransformers on Ubuntu 22.04 is caused by the current system's g++ version being too old, and the pre-defined macros do not include avx_bf16. We have tested and confirmed that it works on g++ 11.4 in Ubuntu 22.04.
### Q: Using fp8 prefill very slow.
The FP8 kernel is build by JIT, so the first run will be slow. The subsequent runs will be faster.
### Q: Possible ways to run graphics cards using volta and turing architectures
From: https://github.com/kvcache-ai/ktransformers/issues/374
1. First, download the latest source code using git.
2. Then, modify the DeepSeek-V3-Chat-multi-gpu-4.yaml in the source code and all related yaml files, replacing all instances of KLinearMarlin with KLinearTorch.
3. Next, you need to compile from the ktransformer source code until it successfully compiles on your local machine.
4. Then, install flash-attn. It won't be used, but not installing it will cause an error.
5. Then, modify local_chat.py, replacing all instances of flash_attention_2 with eager.
6. Then, run local_chat.py. Be sure to follow the official tutorial's commands and adjust according to your local machine's parameters.
7. During the running process, check the memory usage. Observe its invocation through the top command. The memory capacity on a single CPU must be greater than the complete size of the model. (For multiple CPUs, it's just a copy.)
Finally, confirm that the model is fully loaded into memory and specific weight layers are fully loaded into the GPU memory. Then, try to input content in the chat interface and observe if there are any errors.
Attention, for better perfomance, you can check this [method](https://github.com/kvcache-ai/ktransformers/issues/374#issuecomment-2667520838) in the issue
>
>https://github.com/kvcache-ai/ktransformers/blob/89f8218a2ab7ff82fa54dbfe30df741c574317fc/ktransformers/operators/attention.py#L274-L279
>
>```diff
>+ original_dtype = query_states.dtype
>+ target_dtype = torch.half
>+ query_states = query_states.to(target_dtype)
>+ compressed_kv_with_k_pe = compressed_kv_with_k_pe.to(target_dtype)
>+ compressed_kv = compressed_kv.to(target_dtype)
>+ attn_output = attn_output.to(target_dtype)
>
>decode_attention_fwd_grouped(query_states, compressed_kv_with_k_pe, compressed_kv, attn_output,
> page_table,
> position_ids.squeeze(0).to(torch.int32)+1, attn_logits,
> 4, #num_kv_splits # follow vLLM, fix it TODO
> self.softmax_scale,
> past_key_value.page_size)
>
>+ attn_output = attn_output.to(original_dtype)
>```
>
>https://github.com/kvcache-ai/ktransformers/blob/89f8218a2ab7ff82fa54dbfe30df741c574317fc/ktransformers/operators/attention.py#L320-L326
>
>```diff
>- attn_output = flash_attn_func(
>- query_states,
>- key_states,
>- value_states_padded,
>- softmax_scale=self.softmax_scale,
>- causal=True,
>- )
>+ attn_output = F.scaled_dot_product_attention(
>+ query_states.transpose(1, 2),
>+ key_states.transpose(1, 2),
>+ value_states_padded.transpose(1, 2),
>+ scale=self.softmax_scale,
>+ is_causal=True
>+ ).transpose(1, 2)
>```
\ No newline at end of file
doc/en/V3-success.md 0 → 100644
View file @ c009512a
## Hello everyone, here is the successfully reproduced environment configuration for your reference:
### Case 1
- Configuration: l40s 48G + 9654 x2 (192 cores) + 768G DDR5 12-channel
- Performance: prefill 108 tokens/s, decode 10.8 tokens/s
- Used version: main source code compiled
### Case 2
- Configuration: Dual Xeon 6430 32C processors, totaling 64 cores and 128 threads, 480GB DDR5 memory, single 4090 24G graphics card
- Performance: Running speed approximately 6-8 tokens per second
## NOTE
If there are any other configurations that have been successfully run, please feel free to let us know. We will keep updating for everyone to refer to when reproducing. (It has been found that it also works on 2080, AMD, etc. (doge : )
[click here](https://docs.qq.com/smartsheet/form/AVxgQOYhhNfl%2FBB08J2%2Fv3rnnq?tab=BB08J2)
\ No newline at end of file
doc/en/api/server/website.md
View file @ c009512a
...@@ -8,6 +8,20 @@ This document provides the necessary steps to set up and run the web service for ...@@ -8,6 +8,20 @@ This document provides the necessary steps to set up and run the web service for
Before you can compile the web code, make sure you have installed [Node.js](https://nodejs.org) version 18.3 or higher Before you can compile the web code, make sure you have installed [Node.js](https://nodejs.org) version 18.3 or higher
Note: The version of Node.js in the Ubuntu or Debian GNU/Linux software repository is too low, causing compilation errors. Users can also install Node.js through the Nodesource repository, provided they uninstall the outdated version first.
```bash
# sudo apt-get remove nodejs npm -y && sudo apt-get autoremove -y
sudo apt-get update -y && sudo apt-get install -y apt-transport-https ca-certificates curl gnupg
curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key | sudo gpg --dearmor -o /usr/share/keyrings/nodesource.gpg
sudo chmod 644 /usr/share/keyrings/nodesource.gpg
echo "deb [arch=$(dpkg --print-architecture) signed-by=/usr/share/keyrings/nodesource.gpg] https://deb.nodesource.com/node_23.x nodistro main" | sudo tee /etc/apt/sources.list.d/nodesource.list
sudo apt-get update -y
sudo apt-get install nodejs -y
```
Once npm is installed, navigate to the `ktransformers/website` directory: Once npm is installed, navigate to the `ktransformers/website` directory:
```bash ```bash
......
doc/en/benchmark.md 0 → 100644
View file @ c009512a
## Benchmark
To conduct a quick and convenient check, we have employed a simple Python script available [here](https://github.com/kvcache-ai/ktransformers/tree/main/ktransformers/tests) to assess the precision of our **[ktransformers](https://github.com/kvcache-ai/ktransformers)** project. For this evaluation, we utilized the same dataset, which was shuffled in a consistent manner and limited to the first 1,000 data points, to test our implementation across a variety of CPU kernels, MLA kernels, and quantization formats.
We selected the DeepSeek-V3 model in its bf16, int8, and q4km versions for this test. The MMLU dataset, which can be found [here](https://huggingface.co/datasets/cais/mmlu), was used (we selected all datasets and shuffled them with a fixed random seed).
**!!! However, we skipped the few-shot part and only chose the first 1,000 data points for a quick check.** Please note that this approach may result in results that are not consistent with the technical report of DeepSeek-V3. And the test of R1 and further more tests are on going.
To verify our results, we chose [cloud service platform](https://cloud.siliconflow.cn/models) as baseline. All tests were conducted using the same script and datasets, allowing us to make a preliminary assessment of our project's precision.
We set the argument `temperature=0.6`, and to simplify the test process, we skipped the few-shot part and used the following prompt: `There is a single choice question. Answer the question by replying A, B, C, D. No other answers are accepted. Just the letter. \nQuestion: {question}\nA. {option_a}\nB. {option_b}\nC. {option_c}\nD. {option_d}\nAnswer: '`. For more details, please refer to the [script](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/tests/mmlu_test.py).
Given that we have only tested 1,000 cases, which provides only a preliminary judgment, some fluctuations in the results are reasonable. We selected all datasets and shuffled them with a fixed random seed to ensure consistency.
## Some Details
- The bf16 model of DeepSeek-V3 is available [here](https://huggingface.co/opensourcerelease/DeepSeek-V3-bf16/tree/main) (you may convert it to gguf by llama.cpp). The q4km model can be found [here](https://huggingface.co/unsloth/DeepSeek-V3-GGUF/tree/main/DeepSeek-V3-Q4_K_M).
- The optimization YAML file is located [here](https://github.com/kvcache-ai/ktransformers/tree/main/ktransformers/optimize/optimize_rules). For the GEMM Kernel, you can change `KLinearMarlin` to `KLinearTorch`.
- To switch the MLA Kernel from Triton to Torch, you can check and modify [this file](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/operators/attention.py), specifically by using the `forward_windows` method.
- When attempting to conduct the bf16 test (both CPU Weight and GPU Weight), you may encounter issues stemming from older versions of g++ and as, particularly when using Ubuntu 20 or earlier versions. To facilitate a smoother experience and enable you to reproduce our results, we have provided a development container. This container offers a pre-configured environment tailored for this purpose. However, please note that the container does not have the ktrans package installed. Therefore, you may still need to manually install certain packages to ensure everything runs smoothly.
- You may config the model mount dir in `devcontainer/devcontainer.json`, check the `"mouts":` config.
## The Result Table
Uses DeepSeek-V3 model (Some specific cases are R1)
| | | | | | | | |
| ------------------------ | ----------------- | ---------- | ----------------- | ------- | ---------- | ------------------------------------------------------ | ------------ |
| DataSet | CPU Weight Format | CPU Kernel | GPU Weight Format | GEMM Kernel | MLA Kernel | [Siliconflow](https://cloud.siliconflow.cn/models)<br> | Ktrans Point |
| MMLU<br><br>(shuffle 1k) | | | | | | | |
| 1 | bf16 | cpuinfer | bf16 | torch | torch | 81.6 | 81.9 |
| 2 | q8_0 | cpuinfer | bf16 | torch | torch | 81.6 | 83.1 |
| 3 | q4km | cpuinfer | bf16 | torch | triton | 81.6 | 81.4 |
| 4 | q4km | cpuinfer | q4km->marlin 8 | marlin | triton | 81.6 | 81.1 |
| 5 | q4km | cpuinfer | q4km->marlin 4 | marlin | triton | 81.6 | 81 |
| 6 | q4km | cpuinfer | fp8 | fp8gemm | triton | 81.6 | 81.5 |
| 7 (DeepSeek-R1) | iq1 | cpuinfer | fp8 | fp8gemm | triton | 78.6 | 83.6 |
| MMLU-pro<br>(shuffle 1k) | | | | | | | |
| 1 | q4km | cpuinfer | fp8 | fp8gemm | triton | 57.7 | 57.6 |
| 2 | q4km | cpuinfer | q4km->marlin 4 | marlin | triton | 57.7 | 57.5 |
| 3 (DeepSeek-R1) | iq1 | cpuinfer | fp8 | fp8gem | triton | 71.9 | tbd |
| HumanEval | tbd | tbd | tbd | tbd | tbd | tbd | tbd |
| GSM8K | tbd | tbd | tbd | tbd | tbd | tbd | tbd |
**The details for each case are listed below**:
By default, The MLA kernel uses triton in linux and torch in windows. But we need to test torch in linux, so we manually modify the [file](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/operators/attention.py#L592). Just get rid of all the if branch and force it to use `self.forward_windows`
- MMLU test
1. [v3-chat_yaml](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat.yaml) change all the `KLinearMarlin` to `KLinearTorch` (just find all the usage in this file). The source weight comes from [there](https://huggingface.co/opensourcerelease/DeepSeek-V3-bf16) (you need to use llama.cpp to convert it to gguf)
2. [v3-chat_yaml](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat.yaml). You need to modify the code to separately load cpu's expert weight. We leave this as comment in these places: [1](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/operators/experts.py#L122), [2](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/operators/experts.py#L136), [3](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/operators/experts.py#L137) (note in 3, change the path to your local weight file path). The weight file for q8_0 is [here](https://huggingface.co/unsloth/DeepSeek-V3-GGUF/tree/main/DeepSeek-V3-Q8_0)
3. [v3-chat_yaml](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat.yaml). You need to modify the code to separately load cpu's expert weight. We leave this as comment in these places: [1](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/operators/experts.py#L122), [2](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/operators/experts.py#L136), [3](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/operators/experts.py#L137) (note in 3, change the path to your local weight file path). The weight file for q4km is [here](https://huggingface.co/unsloth/DeepSeek-V3-GGUF/tree/main/DeepSeek-V3-Q4_K_M)
4. [v3-chat_yaml](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat.yaml). You don't need to change the source code as they both use q4km. But note the yaml file [here](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat.yaml#L29) and [here](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat.yaml#L18), below these lines you need to add `num_bits: 8` (in other words: add this kwargs to all that use `KLinearMarlin`). The weight file for q4km is [here](https://huggingface.co/unsloth/DeepSeek-V3-GGUF/tree/main/DeepSeek-V3-Q4_K_M)
5. [v3-chat_yaml](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat.yaml). No need to change yaml, just use the default. The weight file for q4km is [here](https://huggingface.co/unsloth/DeepSeek-V3-GGUF/tree/main/DeepSeek-V3-Q4_K_M)
6. You should check the [doc](./fp8_kernel.md) to learn how to test this case. This is a mixture tensor case.
7. You should check the [doc](./fp8_kernel.md) to learn how to test this case. This is a mixture tensor case.
- MMLU-pro test
1. You should check the [doc](./fp8_kernel.md) to learn how to test this case. This is a mixture tensor case.
2. [v3-chat_yaml](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat.yaml). No need to change yaml, just use the default. The weight file for q4km is [here](https://huggingface.co/unsloth/DeepSeek-V3-GGUF/tree/main/DeepSeek-V3-Q4_K_M)
3. You should check the [doc](./fp8_kernel.md) to learn how to test this case. This is a mixture tensor case.
\ No newline at end of file
doc/en/deepseek-v2-injection.md
View file @ c009512a
# Tutorial: Heterogeneous and Local DeepSeek-V2 Inference # Tutorial: Heterogeneous and Local MoE Inference
DeepSeek-(Code)-V2 is a series of strong mixture-of-experts (MoE) models, featuring a total of 236 billion parameters, with 21 billion parameters activated per token. This model has demonstrated remarkable reasoning capabilities across various benchmarks, positioning it as one of the SOTA open models and nearly comparable in performance to GPT-4. DeepSeek-(Code)-V2 is a series of strong mixture-of-experts (MoE) models, featuring a total of 236 billion parameters, with 21 billion parameters activated per token. This model has demonstrated remarkable reasoning capabilities across various benchmarks, positioning it as one of the SOTA open models and nearly comparable in performance to GPT-4. DeepSeek-R1 uses a similar architecture to DeepSeek-V2, but with a bigger number of parameters.
<p align="center"> <p align="center">
<picture> <picture>
...@@ -24,7 +24,7 @@ The following figure provides a brief overview of DeepSeek-V2 architecture. At t ...@@ -24,7 +24,7 @@ The following figure provides a brief overview of DeepSeek-V2 architecture. At t
<p align="center"> <p align="center">
<picture> <picture>
<img alt="DeepSeek on KTransformers" src="../assets/DeepSeek-on-KTransformers.PNG" width=80%> <img alt="DeepSeek on KTransformers" src="../assets/DeepSeek-on-KTransformers.png" width=80%>
</picture> </picture>
</p> </p>
......
doc/en/fp8_kernel.md 0 → 100644
View file @ c009512a
# FP8 Linear Kernel for DeepSeek-V3/R1
## Overview
The DeepSeek-AI team provides FP8 safetensors for DeepSeek-R1/V3 models. We achieve performance optimization through the following works:
- **FP8 GPU Kernel Integration**: FP8 linear layer acceleration kernels integrated in KTransformers
- **Hybrid Quantization Architecture**:
- Attention and Shared-Expert modules use FP8 precision (enhances computational accuracy)
- Experts modules retain GGML quantization (GGUF format, reside in CPU to save GPU memory)
So those who are persuing the best performance can use the FP8 linear kernel for DeepSeek-V3/R1.
## Key Features
✅ Hybrid Precision Architecture (FP8 + GGML)<br>
✅ Memory Optimization (~19GB VRAM usage)
## Quick Start
### Using Pre-Merged Weights
Pre-merged weights are available on Hugging Face:<br>
[KVCache-ai/DeepSeek-V3-GGML-FP8-Hybrid](https://huggingface.co/KVCache-ai/DeepSeek-V3)<br>
[KVCache-ai/DeepSeek-R1-GGML-FP8-Hybrid](https://huggingface.co/KVCache-ai/DeepSeek-R1)
> Please confirm the weights are fully uploaded before downloading. The large file size may extend Hugging Face upload time.
Download Pre-Merged Weights
```shell
pip install -U huggingface_hub
# Optional: Use HF Mirror for faster downloads in special area.
# export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download --resume-download KVCache-ai/DeepSeek-V3-GGML-FP8-Hybrid --local-dir <local_dir>
```
### Using merge scripts
If you got local DeepSeek-R1/V3 fp8 safetensors and gguf weights(eg.q4km), you can merge them using the following scripts.
```shell
python merge_tensors/merge_safetensor_gguf.py \
--safetensor_path <fp8_safetensor_path> \
--gguf_path <gguf_folder_path> \
--output_path <merged_output_path>
```
* `--safetensor_path`: input path of safetensor file([Download](https://huggingface.co/deepseek-ai/DeepSeek-V3/tree/main)).
* `--gguf_path`: input path of gguf folder ([Download](https://huggingface.co/unsloth/DeepSeek-V3-GGUF/tree/main/DeepSeek-V3-Q4_K_M)).
* `--output_path`: output path of merged file.
### Execution Notes
Launch local_chat.py with custom quantized experts
```shell
python ktransformers/local_chat.py \
--model_path deepseek-ai/DeepSeek-V3 \
--gguf_path <merged_weights_folder> \
--optimize_config_path ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-fp8-linear-ggml-experts.yaml \
--cpu_infer <cpu_cores + 1>
```
## Notes
⚠️ Hardware Requirements<br>
* Recommended minimum 19GB available VRAM for FP8 kernel.
* Requires GPU with FP8 support (e.g., 4090)
⏳ First-Run Optimization
JIT compilation causes longer initial execution (subsequent runs retain optimized speed).
🔄 Temporary Interface<br>
Current weight loading implementation is provisional - will be refined in future versions
📁 Path Specification<br>
Despite hybrid quantization, merged weights are stored as .safetensors - pass the containing folder path to `--gguf_path`
\ No newline at end of file
doc/en/injection_tutorial.md
View file @ c009512a
...@@ -59,6 +59,7 @@ Supported operators and their corresponding classes are as follows: ...@@ -59,6 +59,7 @@ Supported operators and their corresponding classes are as follows:
| Linear | KTransformersLinear | KLinearMarlin | Marlin as backend | | Linear | KTransformersLinear | KLinearMarlin | Marlin as backend |
| | | KLinearTorch | pytorch as backend | | | | KLinearTorch | pytorch as backend |
| | | KLinearCPUInfer | llamafile as backend | | | | KLinearCPUInfer | llamafile as backend |
| | | KLinearFP8 | Triton fp8_gemm kernel. Requires GPU be able to caluculate fp8 data |
| experts | KTransformersExperts | KExpertsTorch | pytorch as backend | | experts | KTransformersExperts | KExpertsTorch | pytorch as backend |
| | | KExpertsMarlin | Marlin as backend | | | | KExpertsMarlin | Marlin as backend |
| | | KExpertsCPU | llamafile as backend | | | | KExpertsCPU | llamafile as backend |
......
doc/en/install.md 0 → 100644
View file @ c009512a
<!-- omit in toc -->
# How to Run DeepSeek-R1
- [Preparation](#preparation)
- [Installation](#installation)
- [Attention](#attention)
- [Supported models include:](#supported-models-include)
- [Support quantize format:](#support-quantize-format)
In this document, we will show you how to install and run KTransformers on your local machine. There are two versions:
* V0.2 is the current main branch.
* V0.3 is a preview version only provides binary distribution for now.
* To reproduce our DeepSeek-R1/V3 results, please refer to [Deepseek-R1/V3 Tutorial](./DeepseekR1_V3_tutorial.md) for more detail settings after installation.
## Preparation
Some preparation:
- CUDA 12.1 and above, if you didn't have it yet, you may install from [here](https://developer.nvidia.com/cuda-downloads).
```sh
# Adding CUDA to PATH
if [ -d "/usr/local/cuda/bin" ]; then
export PATH=$PATH:/usr/local/cuda/bin
fi
if [ -d "/usr/local/cuda/lib64" ]; then
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64
# Or you can add it to /etc/ld.so.conf and run ldconfig as root:
# echo "/usr/local/cuda-12.x/lib64" | sudo tee -a /etc/ld.so.conf
# sudo ldconfig
fi
if [ -d "/usr/local/cuda" ]; then
export CUDA_PATH=$CUDA_PATH:/usr/local/cuda
fi
```
- Linux-x86_64 with gcc, g++ and cmake (using Ubuntu as an example)
```sh
sudo apt-get update
sudo apt-get install build-essential cmake ninja-build
```
- We recommend using [Miniconda3](https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh) or [Anaconda3](https://repo.anaconda.com/archive/Anaconda3-2024.10-1-Linux-x86_64.sh) to create a virtual environment with Python=3.11 to run our program. Assuming your Anaconda installation directory is `~/anaconda3`, you should ensure that the version identifier of the GNU C++standard library used by Anaconda includes `GLIBCXX-3.4.32`
```sh
conda create --name ktransformers python=3.11
conda activate ktransformers # you may need to run ‘conda init’ and reopen shell first
conda install -c conda-forge libstdcxx-ng # Anaconda provides a package called `libstdcxx-ng` that includes a newer version of `libstdc++`, which can be installed via `conda-forge`.
strings ~/anaconda3/envs/ktransformers-0.3/lib/libstdc++.so.6 | grep GLIBCXX
```
- Make sure that PyTorch, packaging, ninja is installed You can also [install previous versions of PyTorch](https://pytorch.org/get-started/previous-versions/)
```
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
pip3 install packaging ninja cpufeature numpy
```
- At the same time, you should download and install the corresponding version of flash-attention from https://github.com/Dao-AILab/flash-attention/releases.
## Installation
### Attention
If you want to use numa support, not only do you need to set USE_NUMA=1, but you also need to make sure you have installed the libnuma-dev (`sudo apt-get install libnuma-dev` may help you).
<!-- 1. ~~Use a Docker image, see [documentation for Docker](./doc/en/Docker.md)~~
>We are working on the latest docker image, please wait for a while.
2. ~~You can install using Pypi (for linux):~~
> We are working on the latest pypi package, please wait for a while.
```
pip install ktransformers --no-build-isolation
```
for windows we prepare a pre compiled whl package on [ktransformers-0.2.0+cu125torch24avx2-cp312-cp312-win_amd64.whl](https://github.com/kvcache-ai/ktransformers/releases/download/v0.2.0/ktransformers-0.2.0+cu125torch24avx2-cp312-cp312-win_amd64.whl), which require cuda-12.5, torch-2.4, python-3.11, more pre compiled package are being produced. -->
* Download source code and compile:
- init source code
```sh
git clone https://github.com/kvcache-ai/ktransformers.git
cd ktransformers
git submodule init
git submodule update
```
- [Optional] If you want to run with website, please [compile the website](./api/server/website.md) before execute ```bash install.sh```
- For Linux
- For simple install:
```shell
bash install.sh
```
- For those who have two cpu and 1T RAM:
```shell
# Make sure your system has dual sockets and double size RAM than the model's size (e.g. 1T RAM for 512G model)
apt install libnuma-dev
export USE_NUMA=1
bash install.sh # or #make dev_install
```
- For Windows
```shell
install.bat
```
* If you are developer, you can make use of the makefile to compile and format the code. <br> the detailed usage of makefile is [here](./makefile_usage.md)
<h3>Local Chat</h3>
We provide a simple command-line local chat Python script that you can run for testing.
> Note: this is a very simple test tool only support one round chat without any memory about last input, if you want to try full ability of the model, you may go to [RESTful API and Web UI](#id_666).
<h4>Run Example</h4>
```shell
# Begin from root of your cloned repo!
# Begin from root of your cloned repo!!
# Begin from root of your cloned repo!!!
# Download mzwing/DeepSeek-V2-Lite-Chat-GGUF from huggingface
mkdir DeepSeek-V2-Lite-Chat-GGUF
cd DeepSeek-V2-Lite-Chat-GGUF
wget https://huggingface.co/mradermacher/DeepSeek-V2-Lite-GGUF/resolve/main/DeepSeek-V2-Lite.Q4_K_M.gguf -O DeepSeek-V2-Lite-Chat.Q4_K_M.gguf
cd .. # Move to repo's root dir
# Start local chat
python -m ktransformers.local_chat --model_path deepseek-ai/DeepSeek-V2-Lite-Chat --gguf_path ./DeepSeek-V2-Lite-Chat-GGUF
# If you see “OSError: We couldn't connect to 'https://huggingface.co' to load this file”, try：
# GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite
# python ktransformers.local_chat --model_path ./DeepSeek-V2-Lite --gguf_path ./DeepSeek-V2-Lite-Chat-GGUF
```
It features the following arguments:
- `--model_path` (required): Name of the model (such as "deepseek-ai/DeepSeek-V2-Lite-Chat" which will automatically download configs from [Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite)). Or if you already got local files you may directly use that path to initialize the model.
> Note: <strong>.safetensors</strong> files are not required in the directory. We only need config files to build model and tokenizer.
- `--gguf_path` (required): Path of a directory containing GGUF files which could that can be downloaded from [Hugging Face](https://huggingface.co/mzwing/DeepSeek-V2-Lite-Chat-GGUF/tree/main). Note that the directory should only contains GGUF of current model, which means you need one separate directory for each model.
- `--optimize_config_path` (required except for Qwen2Moe and DeepSeek-V2): Path of YAML file containing optimize rules. There are two rule files pre-written in the [ktransformers/optimize/optimize_rules](ktransformers/optimize/optimize_rules) directory for optimizing DeepSeek-V2 and Qwen2-57B-A14, two SOTA MoE models.
- `--max_new_tokens`: Int (default=1000). Maximum number of new tokens to generate.
- `--cpu_infer`: Int (default=10). The number of CPUs used for inference. Should ideally be set to the (total number of cores - 2).
<details>
<summary>Supported Models/quantization</summary>
### Supported models include:
| ✅ **Supported Models** | ❌ **Deprecated Models** |
|------------------------|------------------------|
| DeepSeek-R1 | ~~InternLM2.5-7B-Chat-1M~~ |
| DeepSeek-V3 | |
| DeepSeek-V2 | |
| DeepSeek-V2.5 | |
| Qwen2-57B | |
| DeepSeek-V2-Lite | |
| Mixtral-8x7B | |
| Mixtral-8x22B | |
### Support quantize format:
| ✅ **Supported Formats** | ❌ **Deprecated Formats** |
|--------------------------|--------------------------|
| Q2_K_L | ~~IQ2_XXS~~ |
| Q2_K_XS | |
| Q3_K_M | |
| Q4_K_M | |
| Q5_K_M | |
| Q6_K | |
| Q8_0 | |
</details>
<details>
<summary>Suggested Model</summary>
| Model Name | Model Size | VRAM | Minimum DRAM | Recommended DRAM |
| ------------------------------ | ---------- | ----- | --------------- | ----------------- |
| DeepSeek-R1-q4_k_m | 377G | 14G | 382G | 512G |
| DeepSeek-V3-q4_k_m | 377G | 14G | 382G | 512G |
| DeepSeek-V2-q4_k_m | 133G | 11G | 136G | 192G |
| DeepSeek-V2.5-q4_k_m | 133G | 11G | 136G | 192G |
| DeepSeek-V2.5-IQ4_XS | 117G | 10G | 107G | 128G |
| Qwen2-57B-A14B-Instruct-q4_k_m | 33G | 8G | 34G | 64G |
| DeepSeek-V2-Lite-q4_k_m | 9.7G | 3G | 13G | 16G |
| Mixtral-8x7B-q4_k_m | 25G | 1.6G | 51G | 64G |
| Mixtral-8x22B-q4_k_m | 80G | 4G | 86.1G | 96G |
| InternLM2.5-7B-Chat-1M | 15.5G | 15.5G | 8G(32K context) | 150G (1M context) |
More will come soon. Please let us know which models you are most interested in.
Be aware that you need to be subject to their corresponding model licenses when using [DeepSeek](https://huggingface.co/deepseek-ai/DeepSeek-V2/blob/main/LICENSE) and [QWen](https://huggingface.co/Qwen/Qwen2-72B-Instruct/blob/main/LICENSE).
</details>
<details>
<summary>Click To Show how to run other examples</summary>
* Qwen2-57B
```sh
pip install flash_attn # For Qwen2
mkdir Qwen2-57B-GGUF && cd Qwen2-57B-GGUF
wget https://huggingface.co/Qwen/Qwen2-57B-A14B-Instruct-GGUF/resolve/main/qwen2-57b-a14b-instruct-q4_k_m.gguf?download=true -O qwen2-57b-a14b-instruct-q4_k_m.gguf
cd ..
python -m ktransformers.local_chat --model_name Qwen/Qwen2-57B-A14B-Instruct --gguf_path ./Qwen2-57B-GGUF
# If you see “OSError: We couldn't connect to 'https://huggingface.co' to load this file”, try：
# GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Qwen/Qwen2-57B-A14B-Instruct
# python ktransformers/local_chat.py --model_path ./Qwen2-57B-A14B-Instruct --gguf_path ./DeepSeek-V2-Lite-Chat-GGUF
```
* Deepseek-V2
```sh
mkdir DeepSeek-V2-Chat-0628-GGUF && cd DeepSeek-V2-Chat-0628-GGUF
# Download weights
wget https://huggingface.co/bartowski/DeepSeek-V2-Chat-0628-GGUF/resolve/main/DeepSeek-V2-Chat-0628-Q4_K_M/DeepSeek-V2-Chat-0628-Q4_K_M-00001-of-00004.gguf -o DeepSeek-V2-Chat-0628-Q4_K_M-00001-of-00004.gguf
wget https://huggingface.co/bartowski/DeepSeek-V2-Chat-0628-GGUF/resolve/main/DeepSeek-V2-Chat-0628-Q4_K_M/DeepSeek-V2-Chat-0628-Q4_K_M-00002-of-00004.gguf -o DeepSeek-V2-Chat-0628-Q4_K_M-00002-of-00004.gguf
wget https://huggingface.co/bartowski/DeepSeek-V2-Chat-0628-GGUF/resolve/main/DeepSeek-V2-Chat-0628-Q4_K_M/DeepSeek-V2-Chat-0628-Q4_K_M-00003-of-00004.gguf -o DeepSeek-V2-Chat-0628-Q4_K_M-00003-of-00004.gguf
wget https://huggingface.co/bartowski/DeepSeek-V2-Chat-0628-GGUF/resolve/main/DeepSeek-V2-Chat-0628-Q4_K_M/DeepSeek-V2-Chat-0628-Q4_K_M-00004-of-00004.gguf -o DeepSeek-V2-Chat-0628-Q4_K_M-00004-of-00004.gguf
cd ..
python -m ktransformers.local_chat --model_name deepseek-ai/DeepSeek-V2-Chat-0628 --gguf_path ./DeepSeek-V2-Chat-0628-GGUF
# If you see “OSError: We couldn't connect to 'https://huggingface.co' to load this file”, try：
# GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/deepseek-ai/DeepSeek-V2-Chat-0628
# python -m ktransformers.local_chat --model_path ./DeepSeek-V2-Chat-0628 --gguf_path ./DeepSeek-V2-Chat-0628-GGUF
```
| model name | weights download link |
|----------|----------|
| Qwen2-57B | [Qwen2-57B-A14B-gguf-Q4K-M](https://huggingface.co/Qwen/Qwen2-57B-A14B-Instruct-GGUF/tree/main) |
| DeepseekV2-coder |[DeepSeek-Coder-V2-Instruct-gguf-Q4K-M](https://huggingface.co/LoneStriker/DeepSeek-Coder-V2-Instruct-GGUF/tree/main) |
| DeepseekV2-chat |[DeepSeek-V2-Chat-gguf-Q4K-M](https://huggingface.co/bullerwins/DeepSeek-V2-Chat-0628-GGUF/tree/main) |
| DeepseekV2-lite | [DeepSeek-V2-Lite-Chat-GGUF-Q4K-M](https://huggingface.co/mzwing/DeepSeek-V2-Lite-Chat-GGUF/tree/main) |
| DeepSeek-R1 | [DeepSeek-R1-gguf-Q4K-M](https://huggingface.co/unsloth/DeepSeek-R1-GGUF/tree/main/DeepSeek-R1-Q4_K_M) |
</details>
<!-- pin block for jump -->
<span id='id_666'>
<h3>RESTful API and Web UI </h3>
Start without website:
```sh
ktransformers --model_path deepseek-ai/DeepSeek-V2-Lite-Chat --gguf_path /path/to/DeepSeek-V2-Lite-Chat-GGUF --port 10002
```
Start with website:
```sh
ktransformers --model_path deepseek-ai/DeepSeek-V2-Lite-Chat --gguf_path /path/to/DeepSeek-V2-Lite-Chat-GGUF --port 10002 --web True
```
Or you want to start server with transformers, the model_path should include safetensors
```bash
ktransformers --type transformers --model_path /mnt/data/model/Qwen2-0.5B-Instruct --port 10002 --web True
```
Access website with url [http://localhost:10002/web/index.html#/chat](http://localhost:10002/web/index.html#/chat) :
<p align="center">
<picture>
<img alt="Web UI" src="https://github.com/user-attachments/assets/615dca9b-a08c-4183-bbd3-ad1362680faf" width=90%>
</picture>
</p>
More information about the RESTful API server can be found [here](doc/en/api/server/server.md). You can also find an example of integrating with Tabby [here](doc/en/api/server/tabby.md).
doc/en/multi-gpu-tutorial.md 0 → 100644
View file @ c009512a
# Muti-GPU
Assume you have read the [Injection Tutorial](./injection_tutorial.md) and have a basic understanding of how to inject a model. In this tutorial, we will show you how to use KTransformers to run a model on multiple GPUs.
If you have multiple GPUs, you can set the device for each module to different GPUs.
DeepseekV2-Chat got 60 layers, if we got 2 GPUs, we can allocate 30 layers to each GPU. Complete multi GPU rule examples [here](https://github.com/kvcache-ai/ktransformers/blob/main/ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-multi-gpu.yaml).
<p align="center">
<picture>
<img alt="Inject-Struction" src="../assets/multi_gpu.png" width=60%>
</picture>
</p>
First of all, for multi-GPU, we have to inject an new operator `KDeepseekV2Model`. And set division of the layers to different GPUs. For our case, we have to set the `transfer_map` in the `KDeepseekV2Model` operatoras as follows:
```yaml
- match:
name: "^model$"
replace:
class: "ktransformers.operators.models.KDeepseekV2Model"
kwargs:
transfer_map:
30: "cuda:1"
```
And we have to set the device for each module in the model.
For example, for `routed experts`, the yaml for one GPU is:
```yaml
- match:
name: "^model\\.layers\\..*\\.mlp\\.experts$"
replace:
class: ktransformers.operators.experts.KTransformersExperts # Custom MoE kernel with expert parallelism
kwargs:
generate_device: "cuda:0"
generate_op: "MLPCUDAExperts"
out_device: "cuda:0"
recursive: False # Don't recursively inject submodules of this module
```
But for two GPUs, we need to set the device for each module in the model.
```yaml
# allcate 0-29 layers‘s out_device to cuda:0
- match:
name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\.mlp\\.experts$"
replace:
class: ktransformers.operators.experts.KTransformersExperts # custom MoE Kernel with expert paralleism
kwargs:
generate_device: "cpu"
generate_op: "KExpertsCPU"
out_device: "cuda:0"
recursive: False # don't recursively inject submodules of this module
# allocate 30-59 layers‘s out_device to cuda:1
- match:
name: "^model\\.layers\\.([345][0-9])\\.mlp\\.experts$"
replace:
class: ktransformers.operators.experts.KTransformersExperts # custom MoE Kernel with expert paralleism
kwargs:
generate_device: "cpu"
generate_op: "KExpertsCPU"
out_device: "cuda:1"
recursive: False # don't recursively inject submodules of this module
```
For other modules, we can set the device in the same way.
# How to fully utilize multi-GPU's VRAM
When you have multiple GPUs, you can fully utilize the VRAM of each GPU by moving more weights to the GPU.
For example, for DeepSeekV2-Chat, we can move the weights of the experts to the GPU.
For example, the yaml for two GPUs is:
```yaml
- match:
name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\.mlp\\.experts$"
replace:
class: ktransformers.operators.experts.KTransformersExperts
kwargs:
generate_device: "cpu"
generate_op: "KExpertsCPU"
out_device: "cuda:0"
recursive: False
```
But we got extra 60GB VRAM on cuda:0, we can move experts in layer 4~8 to cuda:0.
```yaml
# Add new rule before old rule.
- match:
name: "^model\\.layers\\.([4-8])\\.mlp\\.experts$" # inject experts in layer 4~8 as marlin expert
replace:
class: ktransformers.operators.experts.KTransformersExperts
kwargs:
generate_device: "cuda:0"
generate_op: "KExpertsMarlin"
recursive: False
- match:
name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\.mlp\\.experts$"
replace:
class: ktransformers.operators.experts.KTransformersExperts
kwargs:
generate_device: "cpu"
generate_op: "KExpertsCPU"
out_device: "cuda:0"
recursive: False
```
Adjust the layer range as you want. Note that:
* The loading speed will be significantly slower for each expert moved to the GPU.
* You have to close the cuda graph if you want to move the experts to the GPU.
* For DeepSeek-R1/V3, each expert moved to the GPU will consume approximately 6GB of VRAM.
* The first matched rule in yaml will be applied. For example, if you have two rules that match the same layer, only the first rule's replacement will be valid.
doc/zh/DeepseekR1_V3_tutorial_zh.md 0 → 100644
View file @ c009512a
<!-- omit in toc -->
# GPT-4/o1 级别本地 VSCode Copilot 在仅 24GB 显存的台式机上的表现
- [摘要](#摘要)
- [先决条件](#先决条件)
- [基准测试结果](#基准测试结果)
- [V0.2](#v02)
- [设置](#设置)
- [内存占用](#内存占用)
- [基准测试结果](#基准测试结果)
- [V0.3-Preview](#V0.3-Preview)
- [设置](#设置-1)
- [内存占用](#内存占用-1)
- [基准测试结果](#基准测试结果-1)
- [如何运行](#如何运行)
- [V0.2 展示](#v02-展示)
- [单插槽版本 (32 核心)](#单插槽版本（32 核心）)
- [双插槽版本 (64 核心)](#双插槽版本（64 核心）)
- [V0.3 展示](#v03-展示)
- [双插槽版本 (64 核心)](#双插槽版本（64 核心）-1)
- [一些解释](#一些解释)
- [常见问题解答](#常见问题解答)
- [R1 不思考](#R1 不返回思考过程)
- [更多常见问题解答](#更多常见问题解答)
# 摘要
> **2025年2月10日**: 支持在单个（24GB 显存）/多个 GPU 和 382GB 内存上运行 DeepseekR1 和 V3，速度提升高达 3~28 倍。<br>
嗨，我们是 KTransformers 团队（以前因本地 CPU/GPU 混合推理开源项目 DeepSeek-V2 而闻名）。
我们听到了您对 DeepSeek-R1/V3 支持的请求——我们很高兴终于可以交付了！很抱歉让您久等了，但我们一直在酝酿一些真正令人惊叹的东西！
今天，我们自豪地宣布，我们不仅支持 DeepSeek-R1/V3，如下视频所示：
https://github.com/user-attachments/assets/ebd70bfa-b2c1-4abb-ae3b-296ed38aa285
</p>
- **[NEW!!!] 本地 671B DeepSeek-Coder-V3/R1:** 仅使用 14GB 显存和 382GB 内存运行其 Q4_K_M 版本。
- 预填充(Prefill)速度 (tokens/s):
- KTransformers: 54.21 (32 核心) → 74.362 (双插槽，2×32 核心) → 255.26 (优化的 AMX 基 MoE 内核，仅 V0.3) → 286.55 (选择性使用 6 个专家，仅 V0.3)
- 与 llama.cpp 在 2×32 核心下 10.31 tokens/s 相比，速度提升高达 **27.79 倍**
- 解码(Decode)速度 (tokens/s):
- KTransformers: 8.73 (32 核心) → 11.26 (双插槽， 2×32 核心) → 13.69 (选择性使用 6 个专家，仅 V0.3)
- 与 llama.cpp 在 2×32 核心下 4.51 tokens/s 相比，速度提升高达 **3.03 倍**
我们还提供了即将推出的优化预览，包括英特尔 AMX 加速内核和选择性专家激活方法，这将显著提升性能。通过 V0.3 预览版，我们在预填充方面实现了高达 286 tokens/s 的速度，比本地推理的 llama.cpp **快 28 倍**。二进制发行版现已可用，源代码即将推出！请查看 wheel 包 [此处](https://github.com/kvcache-ai/ktransformers/releases/download/v0.1.4/ktransformers-0.3.0rc0+cu126torch26fancy-cp311-cp311-linux_x86_64.whl) 。
## 先决条件
我们在以下配置下进行了最佳性能测试（V0.2）： <br>
CPU: Intel (R) Xeon (R) Gold 6454S 1T 内存 (2 NUMA 节点) <br>
GPU: 4090D 24G 显存 <br>
内存: 标准 DDR5-4800 服务器内存 (1 TB)
## 基准测试结果
### V0.2
#### 设置
- Model: DeepseekV3-q4km (int4)<br>
- CPU: cpu_model_name: Intel (R) Xeon (R) Gold 6454S，每个插槽 32 核心，2 个插槽，2 个 NUMA 节点
- GPU: 4090D 24G 显存
- 我们在充分预热后进行测试
#### 内存占用:
- 单插槽: 382G 内存，至少 14GB 显存
- 双插槽: 1T 内存，至少 14GB 显存
#### 基准测试结果
“6 个专家” 情况是 V0.3 预览版中内容
| Prompt<br>(500 tokens) | 双插槽 Ktrans (6 个专家) | 双插槽 Ktrans (8 个专家) | Single socket Ktrans (6 个专家) | Single socket Ktrans (8 个专家)| llama.cpp (8 个专家) |
|------------------------| --- | --- | --- | --- | --- |
| 预填充(Prefill) token/s | 97.32 | 82.94 | 65.14 | 54.21 | 10.31 |
| 解码(Decode) token/s | 13.69 | 12.208 | 10.303 | 8.73 |4.51 |
**最高加速比在解码方面达到 <u>3.03x</u> 倍，在预填充方面达到 <u>9.44x</u> 倍。**
### V0.3-Preview
#### 设置
- Model: DeepseekV3-BF16 (在线量化为 CPU 的 int8 和 GPU 的 int4)
- CPU: cpu_model_name: Intel (R) Xeon (R) Gold 6454S，每个插槽 32 核心，2 个插槽，2 个 NUMA 节点
- GPU: (1~4)x 4090D 24G 显存 (更长的 prompt 需要更多显存)
#### 内存占用:
- 644GB 内存，至少 14GB 显存
#### 基准测试结果
| Prompt length | 1K | 2K | 4K | 8K |
|---------------|-----|-----|-----|-----|
| KTrans (8 个专家) Prefill token/s | 185.96 | 255.26 | 252.58 | 195.62 |
| KTrans (6 个专家) Prefill token/s | 203.70 | 286.55 | 271.08 | 207.20 |
**KTrans V0.3 的预填充速度比 KTrans V0.2 快 <u>3.45x</u> 倍，比 llama.cpp 快 <u>27.79x</u> 倍。**
**解码速度与 KTrans V0.2（6 个专家版本）相同，因此省略。**
主要加速来自于
- 英特尔 AMX 指令集和我们专门设计的缓存友好内存布局
- 专家选择策略，根据离线配置文件结果选择更少的专家
*从我们对 DeepSeekV2、DeepSeekV3 和 DeepSeekR1 的研究中，当我们略微减少推理中的激活专家数量时，输出质量没有变化。但解码和预填充的速度加快了，这令人鼓舞。因此，我们的展示利用了这一发现。*
## 如何运行
### V0.2 展示
#### 单插槽版本（32 核心）
我们的 local_chat 测试命令是:
``` shell
git clone https://github.com/kvcache-ai/ktransformers.git
cd ktransformers
git submodule init
git submodule update
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 --max_new_tokens 1000
<当您看到聊天时，按回车键加载文本提示文件>
```
`<your model path>` 可以是本地路径，也可以是在线路径，例如 deepseek-ai/DeepSeek-V3。如果在线连接出现问题，可以尝试使用镜像（hf-mirror.com） <br>
`<your gguf path>` 也可以是在线路径，但由于其体积较大，我们建议您下载并量化模型（注意这是目录路径） <br>
`--max_new_tokens 1000` 是最大输出 token 长度。如果发现答案被截断，可以增加此数字以获得更长的答案（但要注意内存不足问题，增加此数字会降低生成速度）.
<br>
命令 numactl -N 1 -m 1 的目的是避免 NUMA 节点之间的数据传输<br>
注意！如果测试 R1 可能会跳过思考。因此，可以添加参数：`--force_think true`，这在 [常见问题解答](#常见问题解答) 部分中解释。
#### 双插槽版本（64 核心）
在安装之前（使用 install.sh 或 `make dev_install`），请确保设置环境变量 `USE_NUMA=1`，方法是 `export USE_NUMA=1`（如果已经安装，请重新安装并设置此环境变量） <br>
我们的 local_chat 测试命令是：
``` shell
git clone https://github.com/kvcache-ai/ktransformers.git
cd ktransformers
git submodule init
git submodule update
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 --max_new_tokens 1000
<当您看到聊天时，按回车键加载文本提示文件>
```
参数的含义相同。但因为我们使用双插槽，所以将 cpu_infer 设置为 65。
### V0.3 展示
#### 双插槽版本（64 核心）
我们的 local_chat 测试命令是：
``` 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 --max_new_tokens 1000
<当您看到聊天时，按回车键加载文本提示文件>
```
参数的含义与 V0.2 相同。但因为我们使用双插槽，所以将 cpu_infer 设置为 65。
## 一些解释
1. 我们还想进一步利用 Xeon Gold CPU 上的两个 NUMA 节点。为了避免节点之间的数据传输成本，我们在两个节点上 "copy" 了关键矩阵，这会增加内存占用，但会加速预填充和解码过程。但这种方法占用大量内存，加载权重时速度较慢，因此加载时请耐心等待并监控内存使用情况。我们计划优化这一巨大的内存开销。敬请期待。
2. 命令参数 `--cpu_infer 65` 指定使用多少核心（超过物理核心数量是可以的，但并不是越多越好。根据实际核心数量适当降低此值）。<br>
3. 为什么使用 CPU/GPU 混合推理？
DeepSeek 的 MLA 操作符计算密集。虽然全部在 CPU 上运行是可行的，但将繁重的计算任务卸载到 GPU 上能带来巨大的性能提升。
4. 加速来自哪里？
- 专家卸载：与传统的基于层或 KVCache 卸载（如 llama.cpp 中的）不同，我们将专家计算卸载到 CPU，将 MLA/KVCache 卸载到 GPU，与 DeepSeek 的架构完美对齐，实现最佳效率。
- 英特尔 AMX 优化 – 我们的 AMX 加速内核经过精心调优，运行速度是现有 llama.cpp 实现的数倍。我们计划在清理后开源此内核，并考虑向 llama.cpp 上游贡献代码。
5. 为什么选择英特尔 CPU？
英特尔目前是唯一支持 AMX 类似指令的 CPU 供应商，与仅支持 AVX 的替代方案相比，性能显著更好。
## 常见问题解答
### R1 不返回思考过程
注意！如果测试 R1 可能会跳过思考。因此，可以添加参数：`--force_think true`。详细信息在 [常见问题解答](./FAQ.md) 部分中。 <br>
## 问题
* 修复服务器集成功能以实现网络API访问支持
* 修复本地聊天功能仅支持单行提示输入的问题（目前输入换行符(\n)即开始生成提示）
### 更多常见问题解答
[详见](./FAQ.md)
install.sh
View file @ c009512a
...@@ -2,6 +2,8 @@ ...@@ -2,6 +2,8 @@
set -e set -e
# clear build dirs # clear build dirs
rm -rf build
rm -rf *.egg-info
rm -rf ktransformers/ktransformers_ext/build rm -rf ktransformers/ktransformers_ext/build
rm -rf ktransformers/ktransformers_ext/cuda/build rm -rf ktransformers/ktransformers_ext/cuda/build
rm -rf ktransformers/ktransformers_ext/cuda/dist rm -rf ktransformers/ktransformers_ext/cuda/dist
......
ktransformers/__init__.py
View file @ c009512a
...@@ -5,7 +5,7 @@ Description : ...@@ -5,7 +5,7 @@ Description :
Author : kkk1nak0 Author : kkk1nak0
Date : 2024-08-15 07:34:46 Date : 2024-08-15 07:34:46
Version : 1.0.0 Version : 1.0.0
LastEditors : unicornchan LastEditors : chenxl
LastEditTime : 2025-02-10 00:59:53 LastEditTime : 2025-02-15 03:53:02
''' '''
__version__ = "0.2.0" __version__ = "0.2.3.post1"
\ No newline at end of file
ktransformers/ktransformers_ext/CMakeLists.txt
View file @ c009512a
...@@ -30,6 +30,9 @@ if (NOT MSVC) ...@@ -30,6 +30,9 @@ if (NOT MSVC)
option(LLAMA_F16C "llama: enable F16C" OFF) option(LLAMA_F16C "llama: enable F16C" OFF)
endif() endif()
option(LLAMA_AVX512_FANCY_SIMD "llama: enable AVX512-VL, AVX512-BW, AVX512-DQ, AVX512-VNNI" OFF) option(LLAMA_AVX512_FANCY_SIMD "llama: enable AVX512-VL, AVX512-BW, AVX512-DQ, AVX512-VNNI" OFF)
option(KTRANSFORMERS_USE_CUDA "ktransformers: use CUDA" OFF)
option(KTRANSFORMERS_USE_MUSA "ktransformers: use MUSA" OFF)
option(KTRANSFORMERS_USE_ROCM "ktransformers: use ROCM" OFF)
# Architecture specific # Architecture specific
# TODO: probably these flags need to be tweaked on some architectures # TODO: probably these flags need to be tweaked on some architectures
...@@ -173,6 +176,7 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW ...@@ -173,6 +176,7 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
list(APPEND ARCH_FLAGS -mavx512bw) list(APPEND ARCH_FLAGS -mavx512bw)
list(APPEND ARCH_FLAGS -mavx512dq) list(APPEND ARCH_FLAGS -mavx512dq)
list(APPEND ARCH_FLAGS -mavx512vnni) list(APPEND ARCH_FLAGS -mavx512vnni)
list(APPEND ARCH_FLAGS -mavx512vpopcntdq)
endif() endif()
if (LLAMA_AVX512_BF16) if (LLAMA_AVX512_BF16)
list(APPEND ARCH_FLAGS -mavx512bf16) list(APPEND ARCH_FLAGS -mavx512bf16)
...@@ -232,18 +236,40 @@ add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/../../third_party/llama.cpp ${CMAKE ...@@ -232,18 +236,40 @@ add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/../../third_party/llama.cpp ${CMAKE
include_directories(${CMAKE_CURRENT_SOURCE_DIR}/../../third_party) include_directories(${CMAKE_CURRENT_SOURCE_DIR}/../../third_party)
if (WIN32) if (WIN32)
include_directories("$ENV{CUDA_PATH}/include") include_directories("$ENV{CUDA_PATH}/include")
add_compile_definitions(KTRANSFORMERS_USE_CUDA=1)
elseif (UNIX) elseif (UNIX)
find_package(CUDA) if (KTRANSFORMERS_USE_CUDA)
find_package(HIP) find_package(CUDA REQUIRED)
find_package(MUSAToolkit)
if(CUDA_FOUND)
include_directories("${CUDA_INCLUDE_DIRS}") include_directories("${CUDA_INCLUDE_DIRS}")
add_compile_definitions(KTRANSFORMERS_USE_CUDA=1)
endif() endif()
if(HIP_FOUND)
include_directories("${HIP_INCLUDE_DIRS}") if (KTRANSFORMERS_USE_ROCM)
find_package(HIP REQUIRED)
if(HIP_FOUND)
include_directories("${HIP_INCLUDE_DIRS}")
add_compile_definitions(KTRANSFORMERS_USE_ROCM=1)
endif()
endif() endif()
if(MUSAToolkit_FOUND)
include_directories("${MUSA_INCLUDE_DIRS}") if (KTRANSFORMERS_USE_MUSA)
if (NOT EXISTS $ENV{MUSA_PATH})
if (NOT EXISTS /opt/musa)
set(MUSA_PATH /usr/local/musa)
else()
set(MUSA_PATH /opt/musa)
endif()
else()
set(MUSA_PATH $ENV{MUSA_PATH})
endif()
list(APPEND CMAKE_MODULE_PATH "${MUSA_PATH}/cmake")
find_package(MUSAToolkit)
if (MUSAToolkit_FOUND)
message(STATUS "MUSA Toolkit found")
add_compile_definitions(KTRANSFORMERS_USE_MUSA=1)
endif()
endif() endif()
endif() endif()
...@@ -260,22 +286,19 @@ target_link_libraries(${PROJECT_NAME} PRIVATE llama) ...@@ -260,22 +286,19 @@ target_link_libraries(${PROJECT_NAME} PRIVATE llama)
if(WIN32) if(WIN32)
target_link_libraries(${PROJECT_NAME} PRIVATE "$ENV{CUDA_PATH}/lib/x64/cudart.lib")#CUDA::cudart target_link_libraries(${PROJECT_NAME} PRIVATE "$ENV{CUDA_PATH}/lib/x64/cudart.lib")#CUDA::cudart
elseif(UNIX) elseif(UNIX)
if(NOT DEFINED ENV{CUDA_HOME} OR "$ENV{CUDA_HOME}" STREQUAL "") if(KTRANSFORMERS_USE_CUDA)
set(ENV{CUDA_HOME} "/usr/local/cuda") if(NOT DEFINED ENV{CUDA_HOME} OR "$ENV{CUDA_HOME}" STREQUAL "")
endif() set(ENV{CUDA_HOME} "/usr/local/cuda")
if(CUDA_FOUND) endif()
add_compile_definitions(USE_CUDA=1)
target_link_libraries(${PROJECT_NAME} PRIVATE "$ENV{CUDA_HOME}/lib64/libcudart.so") target_link_libraries(${PROJECT_NAME} PRIVATE "$ENV{CUDA_HOME}/lib64/libcudart.so")
message(STATUS "Building for CUDA")
endif() endif()
if(HIP_FOUND) if (KTRANSFORMERS_USE_ROCM)
add_compile_definitions(USE_HIP=1) add_compile_definitions(USE_HIP=1)
target_link_libraries(${PROJECT_NAME} PRIVATE "${ROCM_PATH}/lib/libamdhip64.so") target_link_libraries(${PROJECT_NAME} PRIVATE "${ROCM_PATH}/lib/libamdhip64.so")
message(STATUS "Building for HIP") message(STATUS "Building for HIP")
endif() endif()
if(MUSAToolkit_FOUND) if(KTRANSFORMERS_USE_MUSA)
add_compile_definitions(USE_MUSA=1) target_link_libraries(${PROJECT_NAME} PRIVATE MUSA::musart)
message(STATUS "Building for MUSA")
endif() endif()
endif() endif()
......
ktransformers/ktransformers_ext/cpu_backend/backend.cpp
View file @ c009512a
...@@ -54,7 +54,12 @@ void Backend::do_work_stealing_job(int task_num, ...@@ -54,7 +54,12 @@ void Backend::do_work_stealing_job(int task_num,
init_func_ = init_func; init_func_ = init_func;
compute_func_ = compute_func; compute_func_ = compute_func;
finalize_func_ = finalize_func; finalize_func_ = finalize_func;
#ifdef USE_NUMA
// numa node location will be calculated based on the number of threads
thread_num_ = max_thread_num_;
#else
thread_num_ = std::min(max_thread_num_, task_num); thread_num_ = std::min(max_thread_num_, task_num);
#endif
int base = task_num / thread_num_; int base = task_num / thread_num_;
int remain = task_num % thread_num_; int remain = task_num % thread_num_;
thread_state_[0].end = base + (0 < remain); thread_state_[0].end = base + (0 < remain);
...@@ -146,4 +151,4 @@ void Backend::worker_thread(int thread_id) { ...@@ -146,4 +151,4 @@ void Backend::worker_thread(int thread_id) {
return; return;
} }
} }
} }
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment