Merge pull request #663 from kvcache-ai/develop-0.2.2

[release]  Release 0.2.2rc.

.devcontainer/Dockerfile

+FROM pytorch/pytorch:2.3.1-cuda12.1-cudnn8-devel as compile_server
+WORKDIR /workspace
+ENV CUDA_HOME /usr/local/cuda
+RUN <<EOF
+apt update -y &&  apt install -y  --no-install-recommends \
+    git \
+    wget \
+    vim \
+    gcc \
+    g++ \
+    cmake && 
+rm -rf /var/lib/apt/lists/* &&
+cd ktransformers &&
+pip install ninja pyproject numpy cpufeature &&
+pip install flash-attn &&
+cp /usr/lib/x86_64-linux-gnu/libstdc++.so.6 /opt/conda/lib/
+EOF
+# Set the default shell to bash
+CMD ["/bin/bash"]
\ No newline at end of file

.devcontainer/devcontainer.json

+{
+    "name": "Ktrans Dev Container",
+    "privileged": true,
+    "build": {
+        "dockerfile": "Dockerfile",
+        "context": "..",
+        "args": {
+            "http_proxy": "${env:http_proxy}",
+            "https_proxy": "${env:https_proxy}",
+        }
+    },
+    "runArgs": [
+        "--network=host",
+        "--gpus",
+        "all"
+        // "--gpu all"
+    ],
+    "workspaceFolder": "/workspace",
+    "workspaceMount": "source=${localWorkspaceFolder},target=/workspace,type=bind,consistency=cached",
+    "mounts": [
+        "source=/mnt/data,target=/mnt/incontainer,type=bind,consistency=cached"
+    ],
+    "customizations": {
+        "vscode": {
+            "extensions": [
+            ],
+            "settings": {
+                "terminal.integrated.shell.linux": "/bin/bash",
+                "cmake.configureOnOpen": true,
+                "cmake.generator": "Ninja"
+            }
+        }
+    }
+}
\ No newline at end of file

.gitignore

@@ -19,13 +19,9 @@ ktransformers/server/local_store/
 ktransformers/server_test1.db
 *.patch
 img/
-tmp1.txt
-test_65_300_1536.txt
+tmp*.txt
 test.txt
 book
-ktransformers/tests/mmlu_result_silicon.json
 ktransformers/tests/chat_txt.txt
-mmlu_result_q4km.json
-mmlu_result_q4km.log
-ktransformers/tests/mmlu_result_silicon.log
+mmlu_result*
 ktransformers/ktransformers_ext/cuda_musa/

README.md

@@ -23,7 +23,8 @@ Our vision for KTransformers is to serve as a flexible platform for experimentin
 
 <h2 id="Updates">🔥 Updates</h2>
 
-* **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/).
+* **Feb 25, 2025**: Support [FP8 GPU kernel](./doc/en/fp8_kernel.md) for DeepSeek-V3 and R1; [Longer Context](./doc/en/DeepseekR1_V3_tutorial.md#v022-longer-context).
+* **Feb 15, 2025**: Longer Context (from 4K to 8K for 24GB VRAM) & Slightly Faster Speed （+15%, up to 16 Tokens/s), update [docs](./doc/en/DeepseekR1_V3_tutorial.md) and [online books](https://kvcache-ai.github.io/ktransformers/).
 * **Feb 10, 2025**: Support Deepseek-R1 and V3 on single (24GB VRAM)/multi gpu and 382G DRAM, up to 3~28x speedup. For detailed show case and reproduction tutorial, see [here](./doc/en/DeepseekR1_V3_tutorial.md).
 * **Aug 28, 2024**: Decrease DeepseekV2's required VRAM from 21G to 11G.
 * **Aug 15, 2024**: Update detailed [tutorial](doc/en/injection_tutorial.md) for injection and multi-GPU. 
@@ -125,7 +126,7 @@ To utilize the provided kernels, users only need to create a YAML-based injectio
 ```python
 with torch.device("meta"):
     model = AutoModelForCausalLM.from_config(config, trust_remote_code=True)
-optimize_and_load_gguf(model, optimize_rule_path, gguf_path, config)
+optimize_and_load_gguf(model, optimize_config_path, gguf_path, config)
 ...
 generated = prefill_and_generate(model, tokenizer, input_tensor.cuda(), max_new_tokens=1000)
 ```

README_ZH.md

@@ -21,7 +21,8 @@ KTransformers 是一个以 Python 为中心的灵活框架，其核心是可扩
 
 <h2 id="Updates">🔥 更新</h2>
 
-* **2025 年 2 月 15 日**：KTransformers V0.2.1: 长上下文(从4K到8K，24GB VRAM) & 稍快的速度(+15%)(最快 16 Tokens/s)，文档请参见 [这里](./doc/en/DeepseekR1_V3_tutorial.md) 和 [在线指南](https://kvcache-ai.github.io/ktransformers/) 。
+* **2025 年 2 月 15 日**：为DeepSeek-V3/R1支持[FP8 GPU内核](./doc/en/fp8_kernel.md); 支持更长的上下文([教程](./doc/en/DeepseekR1_V3_tutorial.md#v022-longer-context)).
+* **2025 年 2 月 15 日**：长上下文(从4K到8K，24GB VRAM) & 稍快的速度(+15%)(最快 16 Tokens/s)，文档请参见 [这里](./doc/en/DeepseekR1_V3_tutorial.md) 和 [在线指南](https://kvcache-ai.github.io/ktransformers/) 。
 * **2025 年 2 月 10 日**：支持 Deepseek-R1 和 V3 在单个（24GB VRAM）/多 GPU 和 382G DRAM 上运行，速度提升高达 3~28 倍。详细教程请参见 [这里](./doc/en/DeepseekR1_V3_tutorial.md)。
 * **2024 年 8 月 28 日**：支持 InternLM2.5-7B-Chat-1M 模型下的 1M 上下文，使用 24GB 的 VRAM 和 150GB 的 DRAM。详细教程请参见 [这里](./doc/en/long_context_tutorial.md)。
 * **2024 年 8 月 28 日**：将 DeepseekV2 所需的 VRAM 从 21G 降低到 11G。
@@ -68,11 +69,11 @@ https://github.com/user-attachments/assets/4c6a8a38-05aa-497d-8eb1-3a5b3918429c
 
 </p>
 
-<h3>在仅 24GB VRAM 的桌面上进行 1M 上下文本地推理</h3>
-<p align="center">
-
-https://github.com/user-attachments/assets/a865e5e4-bca3-401e-94b8-af3c080e6c12
+<!-- <h3>在仅 24GB VRAM 的桌面上进行 1M 上下文本地推理</h3>
+<p align="center"> -->
 
+<!-- https://github.com/user-attachments/assets/a865e5e4-bca3-401e-94b8-af3c080e6c12 -->
+<!-- 
 * **1M 上下文 InternLM 2.5 7B**：以全 bf16 精度运行，使用 24GB VRAM 和 150GB DRAM，可在本地桌面设置中实现。在 1M "针在干草堆中" 测试中达到 92.88% 的成功率，在 128K NIAH 测试中达到 100%。
 
 <p align="center">
@@ -89,7 +90,7 @@ https://github.com/user-attachments/assets/a865e5e4-bca3-401e-94b8-af3c080e6c12
 
 * **增强的速度**：使用稀疏注意力，通过 llamafile 内核实现 1M 上下文生成 16.91 tokens/s 的速度。这种方法比 llama.cpp 的全注意力方法快 10 倍以上。
 
-* **灵活的稀疏注意力框架**：提供了一个灵活的块稀疏注意力框架，用于 CPU 卸载解码。与 SnapKV、Quest 和 InfLLm 兼容。更多信息请参见 [这里](./doc/en/long_context_introduction.md)。
+* **灵活的稀疏注意力框架**：提供了一个灵活的块稀疏注意力框架，用于 CPU 卸载解码。与 SnapKV、Quest 和 InfLLm 兼容。更多信息请参见 [这里](./doc/en/long_context_introduction.md)。 -->
 
 <strong>更多高级功能即将推出，敬请期待！</strong>
 
@@ -116,7 +117,7 @@ KTransformers 的核心是一个用户友好的、基于模板的注入框架。
 ```python
 with torch.device("meta"):
     model = AutoModelForCausalLM.from_config(config, trust_remote_code=True)
-optimize_and_load_gguf(model, optimize_rule_path, gguf_path, config)
+optimize_and_load_gguf(model, optimize_config_path, gguf_path, config)
 ...
 generated = prefill_and_generate(model, tokenizer, input_tensor.cuda(), max_new_tokens=1000)
 ```
@@ -151,7 +152,7 @@ YAML 文件中的每个规则都有两部分：`match` 和 `replace`。`match`
 
 <h2 id="ack">致谢和贡献者</h2>
 
-KTransformer 的开发基于 Transformers 提供的灵活和多功能框架。我们还受益于 GGUF/GGML、Llamafile 和 Marlin 等高级内核。我们计划通过向上游贡献我们的修改来回馈社区。
+KTransformer 的开发基于 Transformers 提供的灵活和多功能框架。我们还受益于 GGUF/GGML、Llamafile 、 Marlin、sglang和flashinfer 等高级内核。我们计划通过向上游贡献我们的修改来回馈社区。
 
 KTransformer 由清华大学 <a href="https://madsys.cs.tsinghua.edu.cn/">MADSys group</a> 小组的成员以及 <a href="http://approaching.ai/">Approaching.AI</a> 的成员积极维护和开发。我们欢迎新的贡献者加入我们，使 KTransformer 更快、更易于使用。
 

doc/README.md

@@ -22,6 +22,7 @@ Our vision for KTransformers is to serve as a flexible platform for experimentin
 
 <h2 id="Updates">🔥 Updates</h2>
 
+* **Feb 25, 2025**: Support [FP8 GPU kernel](./doc/en/fp8_kernel.md) for DeepSeek-V3 and R1; [Longer Context](./doc/en/DeepseekR1_V3_tutorial.md#v022-longer-context).
 * **Feb 10, 2025**: Support Deepseek-R1 and V3 on single (24GB VRAM)/multi gpu and 382G DRAM, up to 3~28x speedup. The detailed tutorial is [here](./en/DeepseekR1_V3_tutorial.md).
 * **Aug 28, 2024**: Support 1M context under the InternLM2.5-7B-Chat-1M model, utilizing 24GB of VRAM and 150GB of DRAM. The detailed tutorial is [here](./en/long_context_tutorial.md).
 * **Aug 28, 2024**: Decrease DeepseekV2's required VRAM from 21G to 11G.

doc/SUMMARY.md

@@ -5,10 +5,11 @@
 - [Installation Guide](en/install.md)
 
 # Tutorial 
-- [Deepseek-R1/V3 Show Case](en/DeepseekR1_V3_tutorial.md)
+- [Deepseek-R1/V3 Show Case/Tutorial](en/DeepseekR1_V3_tutorial.md)
 - [Why KTransformers So Fast](en/deepseek-v2-injection.md)
 - [Injection Tutorial](en/injection_tutorial.md)
 - [Multi-GPU Tutorial](en/multi-gpu-tutorial.md)
+- [Use FP8 GPU Kernel](en/fp8_kernel.md)
 # Server
   - [Server](en/api/server/server.md)
   - [Website](en/api/server/website.md)

doc/en/DeepseekR1_V3_tutorial.md

@@ -16,6 +16,7 @@
 			- [Memory consumptions:](#memory-consumptions)
 			- [Benchmark results](#benchmark-results-2)
 	- [How to Run](#how-to-run)
+ 		- [V0.2.2 longer context](#v022-longer-context)
 		- [V0.2 \& V0.2.1 Showcase](#v02--v021-showcase)
 			- [Single socket version (32 cores)](#single-socket-version-32-cores)
 			- [Dual socket version (64 cores)](#dual-socket-version-64-cores)
@@ -154,6 +155,18 @@ 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
+### V0.2.2 longer context
+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).
+```
 ### V0.2 & V0.2.1 Showcase
 #### Single socket version (32 cores)
 Our local_chat test command is:

doc/en/FAQ.md

@@ -45,7 +45,7 @@ from-https://github.com/kvcache-ai/ktransformers/issues/129#issue-2842799552
 
 ### 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.
 
@@ -55,7 +55,7 @@ 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.
 
-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
 
@@ -63,9 +63,37 @@ Make sure you:
 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.
 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
 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)
 
 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.
\ No newline at end of file

doc/en/benchmark.md

+## 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 Detail
+
+- 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 Matrix MUL 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
+
+|                          |                   |            |                   |         |            |                                                        |              |
+| ------------------------ | ----------------- | ---------- | ----------------- | ------- | ---------- | ------------------------------------------------------ | ------------ |
+| DataSet                  | CPU Weight Format | CPU Kernel | GPU Weight Format | GEMM    | MLA Kernel | [Siliconflow](https://cloud.siliconflow.cn/models)<br> | Ktrans Point |
+| MMLU<br><br>(shuffle 1k) | bf16              | cpuinfer   | bf16              | torch   | torch      | 81.6                                                   | 81.9         |
+|                          | int8              | cpuinfer   | bf16              | torch   | torch      | 81.6                                                   | 83.1         |
+|                          | q4km              | cpuinfer   | bf16              | torch   | torch      | 81.6                                                   | 82.8         |
+|                          | q4km              | cpuinfer   | bf16              | torch   | triton     | 81.6                                                   | 81.4         |
+|                          | q4km              | cpuinfer   | q4km->marlin 8    | marlin  | triton     | 81.6                                                   | 81.1         |
+|                          | q4km              | cpuinfer   | q4km->marlin 4    | marlin  | triton     | 81.6                                                   | 81           |
+|                          | q4km              | cpuinfer   | fp8               | marlin  | triton     | 81.6                                                   | 81.5         |
+| MMLU-pro                 | q4km              | cpuinfer   | fp8               | fp8gemm | triton     | 57.7                                                   | 57.6         |
+| MMLU-pro                 | q4km              | cpuinfer   | q4km->marlin 4    | marlin  | triton     | 57.7                                                   | 57.5         |
+| HumanEval                | tbd               | tbd        | tbd               | tbd     | tbd        | tbd                                                    | tbd          |
+| GSM8K                    | tbd               | tbd        | tbd               | tbd     | tbd        | tbd                                                    | tbd          |

doc/en/fp8_kernel.md

+# FP8 Linear Kernel for DeepSeek-V3
+
+## 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)
+✅ Memory Optimization (~19GB VRAM usage)
+
+## Quick Start
+### Using Pre-Merged Weights
+
+Pre-merged weights are available on Hugging Face:
+[KVCache-ai/DeepSeek-V3](https://huggingface.co/KVCache-ai/DeepSeek-V3)
+[KVCache-ai/DeepSeek-R1](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 --local-dir <local_dir>
+```
+### Using merge scripts
+If you got local DeepSeek-R1/V3 fp8 safetensors and q4km gguf weights, you can merge them using the following scripts.
+
+```shell
+python convert_model.py \
+  --safetensor_path <fp8_safetensor_path> \
+  --gguf_path <q4km_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
+* 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
+Current weight loading implementation is provisional - will be refined in future versions
+
+📁 Path Specification
+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

@@ -59,6 +59,7 @@ Supported operators and their corresponding classes are as follows:
 | Linear    | KTransformersLinear    | KLinearMarlin           | Marlin as backend    |
 |           |                        | KLinearTorch            | pytorch 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   |
 |           |                        | KExpertsMarlin          | Marlin as backend    |
 |           |                        | KExpertsCPU             | llamafile as backend |

doc/en/install.md

@@ -141,7 +141,7 @@ It features the following arguments:
 
 - `--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_rule_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.
+- `--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.
 

ktransformers/ktransformers_ext/triton/fp8gemm.py

+# Adopted from https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/main/inference/kernel.py
+from typing import Tuple
+
+import torch
+import triton
+import triton.language as tl
+from triton import Config
+
+
+@triton.jit
+def act_quant_kernel(x_ptr, y_ptr, s_ptr, BLOCK_SIZE: tl.constexpr):
+    """
+    Quantizes the input tensor `x_ptr` and stores the result in `y_ptr` and the scaling factor in `s_ptr`.
+
+    Args:
+        x_ptr (triton.Pointer): Pointer to the input tensor.
+        y_ptr (triton.Pointer): Pointer to the output tensor where quantized values will be stored.
+        s_ptr (triton.Pointer): Pointer to the output tensor where scaling factors will be stored.
+        BLOCK_SIZE (tl.constexpr): The size of the block to be processed by each program instance.
+
+    Returns:
+        None
+    """
+    pid = tl.program_id(axis=0)
+    offs = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    x = tl.load(x_ptr + offs).to(tl.float32)
+    s = tl.max(tl.abs(x)) / 448.
+    y = x / s
+    y = y.to(y_ptr.dtype.element_ty)
+    tl.store(y_ptr + offs, y)
+    tl.store(s_ptr + pid, s)
+
+
+def act_quant(x: torch.Tensor, block_size: int = 128) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Quantizes the input tensor `x` using block-wise quantization.
+
+    Args:
+        x (torch.Tensor): The input tensor to be quantized. Must be contiguous and its last dimension size must be divisible by `block_size`.
+        block_size (int, optional): The size of the blocks to be used for quantization. Default is 128.
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: A tuple containing:
+            - The quantized tensor with dtype `torch.float8_e4m3fn`.
+            - A tensor of scaling factors with dtype `torch.float32`.
+    """
+    assert x.is_contiguous(), 'Input tensor must be contiguous'
+    assert x.size(-1) % block_size == 0, f'Last dimension size must be divisible by block_size (block_size={block_size})'
+    y = torch.empty_like(x, dtype=torch.float8_e4m3fn)
+    s = x.new_empty(*x.size()[:-1], x.size(-1) // block_size, dtype=torch.float32)
+    grid = lambda meta: (triton.cdiv(x.numel(), meta['BLOCK_SIZE']), )
+    act_quant_kernel[grid](x, y, s, BLOCK_SIZE=block_size)
+    return y, s
+
+
+@triton.jit
+def weight_dequant_kernel(x_ptr, s_ptr, y_ptr, M, N, BLOCK_SIZE: tl.constexpr):
+    """
+    Dequantizes weights using the provided scaling factors and stores the result.
+
+    Args:
+        x_ptr (tl.pointer): Pointer to the quantized weights.
+        s_ptr (tl.pointer): Pointer to the scaling factors.
+        y_ptr (tl.pointer): Pointer to the output buffer for dequantized weights.
+        M (int): Number of rows in the weight matrix.
+        N (int): Number of columns in the weight matrix.
+        BLOCK_SIZE (tl.constexpr): Size of the block for tiling.
+
+    Returns:
+        None
+    """
+    pid_m = tl.program_id(axis=0)
+    pid_n = tl.program_id(axis=1)
+    n = tl.cdiv(N, BLOCK_SIZE)
+    offs_m = pid_m * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    offs_n = pid_n * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    offs = offs_m[:, None] * N + offs_n[None, :]
+    mask = (offs_m[:, None] < M) & (offs_n[None, :] < N)
+    x = tl.load(x_ptr + offs, mask=mask).to(tl.float32)
+    s = tl.load(s_ptr + pid_m * n + pid_n)
+    y = x * s
+    tl.store(y_ptr + offs, y, mask=mask)
+
+
+def weight_dequant(x: torch.Tensor, s: torch.Tensor, block_size: int = 128) -> torch.Tensor:
+    """
+    Dequantizes the given weight tensor using the provided scale tensor.
+
+    Args:
+        x (torch.Tensor): The quantized weight tensor of shape (M, N).
+        s (torch.Tensor): The scale tensor of shape (M, N).
+        block_size (int, optional): The block size to use for dequantization. Defaults to 128.
+
+    Returns:
+        torch.Tensor: The dequantized weight tensor of the same shape as `x`.
+
+    Raises:
+        AssertionError: If `x` or `s` are not contiguous or if their dimensions are not 2.
+    """
+    assert x.is_contiguous() and s.is_contiguous(), 'Input tensors must be contiguous'
+    assert x.dim() == 2 and s.dim() == 2, 'Input tensors must have 2 dimensions'
+    M, N = x.size()
+    y = torch.empty_like(x, dtype=torch.get_default_dtype())
+    grid = lambda meta: (triton.cdiv(M, meta['BLOCK_SIZE']), triton.cdiv(N, meta['BLOCK_SIZE']))
+    with torch.cuda.device(x.device):
+        weight_dequant_kernel[grid](x, s, y, M, N, BLOCK_SIZE=block_size)
+    return y
+
+
+fp8_gemm_configs = [
+    Config({'BLOCK_SIZE_M': block_m, 'BLOCK_SIZE_N': block_n, 'BLOCK_SIZE_K': 128}, num_stages=num_stages, num_warps=8)
+    for block_m in [16, 32, 64] for block_n in [32, 64, 128] for num_stages in [3, 4, 5, 6]
+]
+
+@triton.autotune(configs=fp8_gemm_configs, key=['N', 'K'])
+@triton.jit
+def fp8_gemm_kernel(a_ptr, b_ptr, c_ptr,
+                    a_s_ptr, b_s_ptr,
+                    M, N: tl.constexpr, K: tl.constexpr,
+                    BLOCK_SIZE_M: tl.constexpr,
+                    BLOCK_SIZE_N: tl.constexpr,
+                    BLOCK_SIZE_K: tl.constexpr):
+    """
+    Performs a matrix multiplication operation on FP8 matrices with scaling factors.
+
+    Args:
+        a_ptr (tl.tensor): Pointer to the first input matrix A.
+        b_ptr (tl.tensor): Pointer to the second input matrix B.
+        c_ptr (tl.tensor): Pointer to the output matrix C.
+        a_s_ptr (tl.tensor): Pointer to the scaling factors for matrix A.
+        b_s_ptr (tl.tensor): Pointer to the scaling factors for matrix B.
+        M (int): Number of rows in matrix A and C.
+        N (tl.constexpr): Number of columns in matrix B and C.
+        K (tl.constexpr): Number of columns in matrix A and rows in matrix B.
+        BLOCK_SIZE_M (tl.constexpr): Block size for the M dimension.
+        BLOCK_SIZE_N (tl.constexpr): Block size for the N dimension.
+        BLOCK_SIZE_K (tl.constexpr): Block size for the K dimension.
+
+    Returns:
+        None
+    """
+    pid_m = tl.program_id(axis=0)
+    pid_n = tl.program_id(axis=1)
+    k = tl.cdiv(K, BLOCK_SIZE_K)
+    offs_m = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
+    offs_n = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
+    offs_k = tl.arange(0, BLOCK_SIZE_K)
+    a_ptrs = a_ptr + offs_m[:, None] * K + offs_k[None, :]
+    b_ptrs = b_ptr + offs_n[None, :] * K + offs_k[:, None]
+    a_s_ptrs = a_s_ptr + offs_m * k
+    b_s_ptrs = b_s_ptr + (offs_n // BLOCK_SIZE_K) * k
+
+    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+    for i in range(k):
+        a = tl.load(a_ptrs, mask=offs_k[None, :] < K - i * BLOCK_SIZE_K, other=0.0)
+        b = tl.load(b_ptrs, mask=offs_k[:, None] < K - i * BLOCK_SIZE_K, other=0.0)
+        a_s = tl.load(a_s_ptrs)
+        b_s = tl.load(b_s_ptrs)
+        accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :]
+        a_ptrs += BLOCK_SIZE_K
+        b_ptrs += BLOCK_SIZE_K
+        a_s_ptrs += 1
+        b_s_ptrs += 1
+    c = accumulator.to(c_ptr.dtype.element_ty)
+    offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    offs_n = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    c_ptrs = c_ptr + offs_m[:, None] * N + offs_n[None, :]
+    mask = (offs_m[:, None] < M) & (offs_n[None, :] < N)
+    tl.store(c_ptrs, c, mask=mask)
+
+
+def fp8_gemm(a: torch.Tensor, a_s: torch.Tensor, b: torch.Tensor, b_s: torch.Tensor):
+    """
+    Perform a matrix multiplication using FP8 precision.
+
+    Args:
+        a (torch.Tensor): The first input matrix, must be contiguous.
+        a_s (torch.Tensor): The scaling factor for the first input matrix, must be contiguous.
+        b (torch.Tensor): The second input matrix, must be contiguous.
+        b_s (torch.Tensor): The scaling factor for the second input matrix, must be contiguous.
+
+    Returns:
+        torch.Tensor: The result of the matrix multiplication.
+    """
+    assert a.is_contiguous() and b.is_contiguous(), 'Input tensors must be contiguous'
+    assert a_s.is_contiguous() and b_s.is_contiguous(), 'Scaling factor tensors must be contiguous'
+    K = a.size(-1)
+    M = a.numel() // K
+    N = b.size(0)
+    c = a.new_empty(*a.size()[:-1], N, dtype=torch.get_default_dtype())
+    grid = lambda META: (triton.cdiv(M, META['BLOCK_SIZE_M']), triton.cdiv(N, META['BLOCK_SIZE_N']))
+    fp8_gemm_kernel[grid](a, b, c, a_s, b_s, M, N, K)
+    return c
\ No newline at end of file

ktransformers/local_chat.py

@@ -28,7 +28,7 @@ from ktransformers.models.modeling_qwen2_moe import Qwen2MoeForCausalLM
 from ktransformers.models.modeling_deepseek_v3 import DeepseekV3ForCausalLM
 from ktransformers.models.modeling_llama import LlamaForCausalLM
 from ktransformers.models.modeling_mixtral import MixtralForCausalLM
-from ktransformers.util.utils import prefill_and_generate
+from ktransformers.util.utils import prefill_and_generate, get_compute_capability
 from ktransformers.server.config.config import Config
 from ktransformers.operators.flashinfer_wrapper import flashinfer_enabled
 
@@ -54,7 +54,7 @@ default_optimize_rules = {
 
 def local_chat(
     model_path: str | None = None,
-    optimize_rule_path: str = None,
+    optimize_config_path: str = None,
     gguf_path: str | None = None,
     max_new_tokens: int = 300,
     cpu_infer: int = Config().cpu_infer,
@@ -94,12 +94,12 @@ def local_chat(
                 config, trust_remote_code=True, attn_implementation="flash_attention_2"
             )
 
-    if optimize_rule_path is None:
+    if optimize_config_path is None:
         if config.architectures[0] in default_optimize_rules:
             print("using default_optimize_rule for", config.architectures[0])
-            optimize_rule_path = default_optimize_rules[config.architectures[0]]
+            optimize_config_path = default_optimize_rules[config.architectures[0]]
         else:
-            optimize_rule_path = input(
+            optimize_config_path = input(
                 "please input the path of your rule file(yaml file containing optimize rules):"
             )
 
@@ -107,7 +107,7 @@ def local_chat(
         gguf_path = input(
             "please input the path of your gguf file(gguf file in the dir containing input gguf file must all belong to current model):"
         )
-    optimize_and_load_gguf(model, optimize_rule_path, gguf_path, config)
+    optimize_and_load_gguf(model, optimize_config_path, gguf_path, config)
     
     try:
             model.generation_config = GenerationConfig.from_pretrained(model_path)
@@ -168,7 +168,7 @@ def local_chat(
             assert Config().long_context_config['max_seq_len'] > input_tensor.shape[1] + max_new_tokens, \
             "please change max_seq_len in  ~/.ktransformers/config.yaml"
         
-        if system != "Windows" and (config.architectures[0] == "DeepseekV2ForCausalLM" or "DeepseekV3ForCausalLM") and flashinfer_enabled:
+        if system != "Windows" and (config.architectures[0] == "DeepseekV2ForCausalLM" or "DeepseekV3ForCausalLM") and flashinfer_enabled and get_compute_capability() >= 8:
             generated = prefill_and_generate(
                 model, tokenizer, input_tensor.cuda(), max_new_tokens, use_cuda_graph, mode = mode, force_think = force_think,
                 use_flashinfer_mla = True, num_heads = config.num_attention_heads, head_dim_ckv = config.kv_lora_rank, head_dim_kpe = config.qk_rope_head_dim, q_head_dim = config.qk_rope_head_dim + config.qk_nope_head_dim

ktransformers/operators/attention.py

@@ -16,6 +16,7 @@ from ktransformers.models.modeling_deepseek import DeepseekV2Attention, apply_ro
 from typing import Optional, Tuple
 from ktransformers.operators.base_operator import BaseInjectedModule
 from ktransformers.util.custom_gguf import GGUFLoader
+from ktransformers.util.utils import get_compute_capability
 import logging
 from transformers.configuration_utils import PretrainedConfig
 from transformers.cache_utils import Cache
@@ -48,12 +49,14 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
                  prefill_device: str = "cuda",
                  generate_device: str = "cuda",
                  chunck_size: int = 1000,
+                 absorb_for_prefill: bool = False,
                  **kwargs):
         BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs)
         self.orig_module.__init__(orig_module.config,
             orig_module.layer_idx)
         self.chunck_size = chunck_size # TODO, generate chunck_size automatically.
         self.mla_wrapper = None
+        self.absorb_for_prefill = absorb_for_prefill
 
     def get_absorbed(self) -> Tuple[torch.Tensor, torch.Tensor]:
         if not (hasattr(self, 'q_absorb') and hasattr(self, 'out_absorb')):
@@ -242,7 +245,7 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
             q_nope = q_nope.transpose(1, 2) # q_len is 1, no GPU overhead, same below
             q_nope = torch.matmul(q_nope, q_absorb) # batched MM
             q_nope = q_nope.transpose(1, 2)
-            assert q_nope.is_contiguous()
+            #assert q_nope.is_contiguous()
             
             # q_nope [bsz, q_len, self.num_heads, self.kv_lora_rank]
             # q_pe [bsz, q_len, self.num_heads, self.qk_rope_head_dim]
@@ -282,6 +285,7 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
             # out_absorb [self.num_heads, self.v_head_dim, self.kv_lora_rank]
             attn_output = attn_output.transpose(1, 2)
             attn_output = torch.matmul(attn_output, out_absorb.mT)
+            attn_output = attn_output.transpose(1, 2)
             
             attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.v_head_dim)
             attn_output = self.o_proj(attn_output)
@@ -380,7 +384,7 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
         # q_pe [bsz, q_len, self.num_heads, self.qk_rope_head_dim] k_pe [bsz, q_len, 1, self.qk_rope_head_dim]
         
         # decode
-        if q_len == 1:
+        if q_len == 1 or self.absorb_for_prefill:
             if past_key_value is not None:
                 cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
                 compressed_kv_with_k_pe, page_table = past_key_value.update(compressed_kv, k_pe, self.layer_idx, cache_kwargs)
@@ -395,29 +399,42 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
             q_nope = q_nope.transpose(1, 2) # q_len is 1, no GPU overhead, same below
             q_nope = torch.matmul(q_nope, q_absorb) # batched MM
             q_nope = q_nope.transpose(1, 2)
-            assert q_nope.is_contiguous()
+            q_nope = q_nope.contiguous()
+            #assert q_nope.is_contiguous()
             
             # q_nope [bsz, q_len, self.num_heads, self.kv_lora_rank]
             # q_pe [bsz, q_len, self.num_heads, self.qk_rope_head_dim]
-            q_nope.squeeze_(1)
-            q_pe.squeeze_(1)
+            q_nope.squeeze_(0)
+            q_pe.squeeze_(0)
 
             # flash attn doesn't support head_dim bigger than 256, use flashinfer
             if self.mla_wrapper is None:
                 self.mla_wrapper = MLAWrapperSingleton.get_instance(self.device, 1, past_key_value.max_pages, use_cuda_graph = True)
-                if self.mla_wrapper.need_plan:
-                    self.mla_wrapper.need_plan = False
+            if self.mla_wrapper.need_plan:
+                self.mla_wrapper.need_plan = False
+                if q_len == 1:
                     self.mla_wrapper.plan(None,None,None,
-                                      position_ids.squeeze(1)+1,
-                                      self.num_heads,
-                                      self.kv_lora_rank,
-                                      self.qk_rope_head_dim,
-                                      past_key_value.page_size,
-                                      self.softmax_scale,
-                                      q_nope.dtype,
-                                      compressed_kv.dtype)
+                                        position_ids.squeeze(1)+1,
+                                        self.num_heads,
+                                        self.kv_lora_rank,
+                                        self.qk_rope_head_dim,
+                                        past_key_value.page_size,
+                                        self.softmax_scale,
+                                        q_nope.dtype,
+                                        compressed_kv.dtype)
+                else:
+                    qo_indptr = torch.tensor([0, q_len], dtype=torch.int32, device=self.device)
+                    kv_len_arr = torch.tensor([position_ids[0, -1].item()+1], dtype=torch.int32, device=self.device)
+                    self.mla_wrapper.plan(qo_indptr,None,None,
+                                        kv_len_arr,
+                                        self.num_heads,
+                                        self.kv_lora_rank,
+                                        self.qk_rope_head_dim,
+                                        past_key_value.page_size,
+                                        self.softmax_scale,
+                                        q_nope.dtype,
+                                        compressed_kv.dtype)
             attn_output = self.mla_wrapper.run(q_nope, q_pe, compressed_kv, k_pe).view(bsz, q_len, self.num_heads, self.kv_lora_rank)
-
             """
             k = (
                 torch.cat([compressed_kv, k_pe], dim=-1)
@@ -443,10 +460,11 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
             # out_absorb [self.num_heads, self.v_head_dim, self.kv_lora_rank]
             attn_output = attn_output.transpose(1, 2) # [bsz, self.num_heads, q_len, self.kv_lora_rank]
             attn_output = torch.matmul(attn_output, out_absorb.mT) # [bsz, self.num_heads, q_len, self.v_head_dim]
+            attn_output = attn_output.transpose(1, 2).contiguous() # [bsz, q_len, self.num_heads, self.kv_lora_rank]
             
             attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.v_head_dim) # [bsz, q_len, self.num_heads * self.v_head_dim]
             attn_output = self.o_proj(attn_output)
-
+            
             return attn_output, None, past_key_value
         else:
             if past_key_value is not None:
@@ -571,7 +589,8 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
         cache_position: Optional[torch.LongTensor] = None,
         **kwargs,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        if os.name == 'nt':
+        if os.name == 'nt' or get_compute_capability()<8:
+            print("for Windows or GPU before ampere, use forward_windows")
             return self.forward_windows(
                 hidden_states,
                 attention_mask,

ktransformers/operators/experts.py

@@ -245,7 +245,16 @@ class KExpertsCPU(KExpertsBase):
         down_type = None
 
         for key in keys:
-            if key + ".ffn_gate_exps.weight" in self.gguf_loader.tensor_info:
+            if self.gguf_loader.safetensor_loader is not None:
+                # using a temp ugly way to temprary load the tensor
+                gate = self.gguf_loader.safetensor_loader.load_tensor(key + ".ffn_gate_exps.weight").numpy()
+                up = self.gguf_loader.safetensor_loader.load_tensor(key + ".ffn_up_exps.weight").numpy()
+                down = self.gguf_loader.safetensor_loader.load_tensor(key + ".ffn_down_exps.weight").numpy()
+                gate_type = self.gguf_loader.safetensor_loader.load_tensor(key + ".ffn_gate_exps.ggml_type").item()
+                up_type = self.gguf_loader.safetensor_loader.load_tensor(key + ".ffn_up_exps.ggml_type").item()
+                down_type = self.gguf_loader.safetensor_loader.load_tensor(key + ".ffn_down_exps.ggml_type").item()
+            
+            elif key + ".ffn_gate_exps.weight" in self.gguf_loader.tensor_info:
                 gate = self.gguf_loader.get_mmap_tensor(key + ".ffn_gate_exps.weight")
                 up = self.gguf_loader.get_mmap_tensor(key + ".ffn_up_exps.weight")
                 down = self.gguf_loader.get_mmap_tensor(key + ".ffn_down_exps.weight")

ktransformers/operators/flashinfer_wrapper.py

@@ -9,7 +9,7 @@ flashinfer_enabled = False
 
 try:
     import flashinfer
-    flashinfer_enabled = False # disabled now, TODO:use new version of flashinfer and enable
+    flashinfer_enabled = True
     print("found flashinfer")
     
 except ImportError:
@@ -122,7 +122,7 @@ class MLAWrapper():
         if kv_indices is None:
             assert self.max_batch_size == 1
             kv_indices = self.kv_indices_buf
-
+        
         self.wrapper.plan(
             qo_indptr,
             kv_indptr,
@@ -132,14 +132,14 @@ class MLAWrapper():
             head_dim_ckv,
             head_dim_kpe,
             page_size,
-            False, # causal is False for decoding
+            True, # causal
             sm_scale,
             q_data_type,
             kv_data_type,
         )
 
     def run(self, q_nope, q_pe, ckv, k_pe, return_lse = False):
-        return self.wrapper.run(q_nope, q_pe, ckv, k_pe, return_lse)
+        return self.wrapper.run(q_nope, q_pe, ckv, k_pe, return_lse = return_lse)
 
 class MLAWrapperSingleton():
     wrappers:dict = {}
@@ -179,6 +179,24 @@ class MLAWrapperSingleton():
                 sm_scale,
                 q_data_type,
                 kv_data_type,)
+            wrapper.need_plan = False
+            
+    @classmethod
+    def need_plan_all(cls):
+        for device, wrapper in cls.wrappers.items():
+            wrapper.need_plan = True
+        
+    @classmethod
+    def reset_buffer(cls):
+        for device, wrapper in cls.wrappers.items():
+            wrapper.qo_indptr_buf[1] = 1 # assert max_batch_size=1 here.
+            
+    @classmethod
+    def update_buffer(cls, max_pages):
+        for device, wrapper in cls.wrappers.items():
+            wrapper.kv_indptr_buf[1] = max_pages # assert max_batch_size=1 here.
+            wrapper.kv_indices_buf = torch.arange(0, max_pages, dtype=torch.int32, device=device)
+            wrapper.wrapper._kv_indices_buf = wrapper.kv_indices_buf
             
 
 if __name__ == "__main__":
@@ -187,8 +205,9 @@ if __name__ == "__main__":
     page_size = 64
     num_heads = 128
 
-    q_nope = torch.randn((1, num_heads, 512), dtype=torch.bfloat16, device="cuda")
-    q_pe = torch.randn((1, num_heads, 64), dtype=torch.bfloat16, device="cuda")
+    q_len = 10
+    q_nope = torch.randn((q_len, num_heads, 512), dtype=torch.bfloat16, device="cuda")
+    q_pe = torch.randn((q_len, num_heads, 64), dtype=torch.bfloat16, device="cuda")
     ckv = torch.randn((max_pages, page_size, 512), dtype=torch.bfloat16, device="cuda")
     k_pe = torch.randn((max_pages, page_size, 64), dtype=torch.bfloat16, device="cuda")
     
@@ -199,10 +218,10 @@ if __name__ == "__main__":
         max_pages,
     )
     
-    kv_len_arr = torch.tensor([10], dtype=torch.int32, device="cuda")
-
+    kv_len_arr = torch.tensor([q_len], dtype=torch.int32, device="cuda")
+    qo_indptr = torch.tensor([0, q_len], dtype=torch.int32, device="cuda")
     wrapper.plan(
-        None,
+        qo_indptr,
         None,
         None,
         kv_len_arr,
@@ -216,6 +235,7 @@ if __name__ == "__main__":
     )
 
     attn_output = wrapper.run(q_nope, q_pe, ckv, k_pe)
+    print(attn_output.shape)
 
     k = (
         torch.cat([ckv, k_pe], dim=-1)
@@ -235,6 +255,7 @@ if __name__ == "__main__":
         False,
         192 ** (-0.5)
     )
+    print(attn_ref.shape)
     
     torch.testing.assert_close(attn_output, attn_ref, rtol=1e-3, atol=1e-3)
     print("test past")
\ No newline at end of file

ktransformers/operators/gate.py

@@ -67,7 +67,14 @@ class KMoEGateBase(ABC):
 
         for key in keys:
             key = ".".join(key.split(".")[:-1])
-            if key + ".ffn_gate_inp.weight" in self.gguf_loader.tensor_info:
+            if self.gguf_loader.safetensor_loader is not None:
+                targets = [".ffn_gate_inp.weight", ".exp_probs_b.bias"]
+                weight = self.gguf_loader.safetensor_loader.load_tensor(key + ".ffn_gate_inp.weight") 
+                e_score_correction_bias = self.gguf_loader.safetensor_loader.load_tensor(key + ".exp_probs_b.bias")
+                weight_type = weight.dtype
+                e_score_correction_bias_type = e_score_correction_bias.dtype
+                res = {"weight": weight, "e_score_correction_bias": e_score_correction_bias,  "weight_type": weight_type, "e_score_correction_bias_type": e_score_correction_bias_type}
+            elif key + ".ffn_gate_inp.weight" in self.gguf_loader.tensor_info:
                 targets = [".ffn_gate_inp.weight", ".exp_probs_b.bias"]
                 tensors = self.load_multi(key, targets, device=device)
                 weight = tensors[".ffn_gate_inp.weight"]
@@ -116,8 +123,8 @@ class KMoEGate(BaseInjectedModule, KMoEGateBase):
             self.orig_module.e_score_correction_bias = nn.Parameter(w["e_score_correction_bias"])
         else:
             raise ValueError("Invalid weight type")
-        self.orig_module.weight = self.orig_module.weight.to(device)
-        self.orig_module.e_score_correction_bias = self.orig_module.e_score_correction_bias.to(device)
+        self.orig_module.weight = nn.Parameter(self.orig_module.weight.to(device))
+        self.orig_module.e_score_correction_bias = nn.Parameter(self.orig_module.e_score_correction_bias.to(device))
 
     def unload(self):
         if self.weight is not None:

ktransformers/operators/linear.py

@@ -26,6 +26,7 @@ from ktransformers.ktransformers_ext.operators.custom_marlin.quantize.utils.marl
 )
 from ktransformers.operators.base_operator import BaseInjectedModule
 from transformers.configuration_utils import PretrainedConfig
+from ktransformers.ktransformers_ext.triton.fp8gemm import fp8_gemm, act_quant, weight_dequant
 from abc import ABC, abstractmethod
 import sys, os
 sys.path.append(os.path.join(os.path.dirname(__file__), "..", "ktransformers_ext", "build"))
@@ -78,7 +79,13 @@ class KLinearBase(ABC):
             keys = [self.key]
 
         for key in keys:
-            if key + ".weight" in self.gguf_loader.tensor_file_map:
+            if self.gguf_loader.safetensor_loader is not None:
+                # using safetensor_loader
+                tensor = self.gguf_loader.safetensor_loader.load_tensor(key+'.weight')
+                weight_scale_inv = self.gguf_loader.safetensor_loader.load_tensor(key+'.weight_scale_inv')
+                return nn.Parameter(tensor), nn.Parameter(weight_scale_inv)
+                
+            elif key + ".weight" in self.gguf_loader.tensor_file_map:
                 if key + ".bias" in self.gguf_loader.tensor_file_map:
                     tensors = self.load_multi(key, ["weight", "bias"], device=device)
                     tensor = tensors["weight"]
@@ -169,7 +176,61 @@ class KLinearTorch(KLinearBase):
         if self.has_bias:
             self.bias = None
 
-
+class KLinearFP8(KLinearBase):
+    # this kernel requires special handling for weight
+    # Please load the weight file downloaded from KVCache.AI
+    marlin_q_w: torch.Tensor
+    marlin_s: torch.Tensor
+    g_idx: torch.Tensor
+    sort_indices: torch.Tensor
+    has_bias: bool
+    weight: torch.Tensor
+    scale_w: torch.Tensor
+    bias: torch.Tensor
+    def __init__(
+        self,
+        key: str,
+        gguf_loader: GGUFLoader,
+        config: PretrainedConfig,
+        orig_module: nn.Module = None,
+        device: str = "cuda",
+        block_size: int = 128,
+        **kwargs,
+    ):
+        super().__init__(key, gguf_loader, config, orig_module, device, **kwargs)
+        self.has_bias = False
+        self.dtype = torch.get_default_dtype()
+        self.block_size = block_size
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = x.to(self.device)
+        orig_dtype = x.dtype        
+        x_quantized, scale_x = act_quant(x, self.block_size)
+        y = fp8_gemm(x_quantized, scale_x, self.weight, self.weight_scale_inv)
+        return y.to(dtype=orig_dtype)
+    
+    def load(self, w: dict | nn.Parameter | tuple | None = None, device: str|None = None):
+        if device is None: device = self.device
+        if w is None: 
+            w = self.load_weight(device=device) 
+        ### TODO fit weight_inv format
+        if isinstance(w, tuple):
+            self.weight = w[0].to(device)
+            self.weight_scale_inv = w[1].to(device)
+            self.has_bias = False
+        else:
+            raise ValueError("Invalid weight type")
+        self.weight = self.weight.to(device)
+        if self.has_bias:
+            self.bias = self.bias.to(device)
+        
+    def unload(self):
+        if self.weight is not None:
+            self.weight = None
+        if self.has_bias:
+            self.bias = None
+        
+        
 class KLinearMarlin(KLinearBase):
     marlin_q_w: torch.Tensor
     marlin_s: torch.Tensor
@@ -404,7 +465,8 @@ class KLinearCPUInfer(KLinearBase):
 LINEAR_MAP = {
     "KLinearMarlin": KLinearMarlin,
     "KLinearTorch": KLinearTorch,
-    "KLinearCPUInfer": KLinearCPUInfer
+    "KLinearCPUInfer": KLinearCPUInfer,
+    "KLinearFP8": KLinearFP8,
 }
 
 class KTransformersLinear(BaseInjectedModule, KLinearBase):
@@ -440,10 +502,11 @@ class KTransformersLinear(BaseInjectedModule, KLinearBase):
     def forward(self, x):
         if self.mode == InferenceState.PREFILL:
             assert self.prefill_linear is not None, "cpu linear is not initialized"
-            return self.prefill_linear.forward(x)
+            y = self.prefill_linear.forward(x)
         else:
             assert self.generate_linear is not None, "gpu linear is not initialized"
-            return self.generate_linear.forward(x)
+            y = self.generate_linear.forward(x)
+        return y
 
     def load(self, w: dict | nn.Parameter | tuple | None = None, mode: InferenceState = InferenceState.GENERATE):
         if not mode: