Merge pull request #64 from eltociear/patch-1

docs: update long_context_introduction.md

doc/en/long_context_introduction.md

@@ -4,7 +4,7 @@
 
 Training larger models and supporting longer text sequences are currently the two most widely agreed-upon directions toward achieving AGI. After lowering the barrier for local inference with trillion-parameter MoE models, the second showcase scenario for KTransformers is reducing the inference barrier for ultra-long context sequences. Recently, both ChatGLM and InternLM have released open-source models supporting 1M tokens of context. This article will use InternLM2.5-7B-Chat-1M as an example to introduce a method that leverages the sparsity of attention to accelerate long-text inference on heterogeneous CPU/GPU systems.
 
-After optimization, KTransformers has achieved native-precision inference for 128K and even 1M tokens of context on a single 24GB GPU with CPU/DRAM support. In the 128K context scenario, the generation speed is 7.1 times faster than llama.cpp, while also achieving 100% accuary on relatively simple test sets like "needle in haystack" and "passkey". On the more challenging dataset kvretrieval, through flexible framework configurations, we can achieve a **6.22x speedup** during inference while obtaining even higher scores than running the original model directly (**21.2 -> 24.4**). In the 1M context scenario on a single 24GB GPU, KTransformers can similarly achieve a 16 tokens/s inference speed, nearly 10 times faster than llama.cpp under the same conditions, with the "needle in haystack" evaluation score even surpassing the original model (**89.31 -> 92.88**).
+After optimization, KTransformers has achieved native-precision inference for 128K and even 1M tokens of context on a single 24GB GPU with CPU/DRAM support. In the 128K context scenario, the generation speed is 7.1 times faster than llama.cpp, while also achieving 100% accuracy on relatively simple test sets like "needle in haystack" and "passkey". On the more challenging dataset kvretrieval, through flexible framework configurations, we can achieve a **6.22x speedup** during inference while obtaining even higher scores than running the original model directly (**21.2 -> 24.4**). In the 1M context scenario on a single 24GB GPU, KTransformers can similarly achieve a 16 tokens/s inference speed, nearly 10 times faster than llama.cpp under the same conditions, with the "needle in haystack" evaluation score even surpassing the original model (**89.31 -> 92.88**).
 
 Project url: https://github.com/kvcache-ai/ktransformers
 
@@ -313,4 +313,4 @@ For example:
 python local_chat.py --model_path="/data/model/internlm2_5_to_llama_1m"  --gguf_path="/data/model/internlm2_5_to_llama_1m" --max_new_tokens=500 --cpu_infer=10  --use_cuda_graph=True  --mode="long_context" --prompt_file="/path/to/file"
 ```
 
-If you've already specified the input text via the prompt_file, just press Enter when the terminal displays chat: to begin.
+If you've already specified the input text via the prompt_file, just press Enter when the terminal displays chat: to begin.