Update app.py

app.py

@@ -5,9 +5,9 @@ from huggingface_hub import hf_hub_download
 from pynvml import *
 nvmlInit()
 gpu_h = nvmlDeviceGetHandleByIndex(0)
-ctx_limit = 512
-title = "RWKV-4 14B fp16"
-desc = f'''DEMO limited to ctxlen {ctx_limit}, and slow because A10g does not have enough VRAM for 14B fp16 (some layers are computed on CPU instead). Links:
+ctx_limit = 1024
+title = "RWKV-4-Pile-14B-20230228-ctx4096-test663"
+desc = f'''DEMO limited to ctxlen {ctx_limit}. Links:
 <a href='https://github.com/BlinkDL/ChatRWKV' target="_blank" style="margin:0 0.5em">ChatRWKV</a>
 <a href='https://github.com/BlinkDL/RWKV-LM' target="_blank" style="margin:0 0.5em">RWKV-LM</a>
 <a href="https://pypi.org/project/rwkv/" target="_blank" style="margin:0 0.5em">RWKV pip package</a>
@@ -18,9 +18,7 @@ os.environ["RWKV_CUDA_ON"] = '1' # if '1' then use CUDA kernel for seq mode (muc
 
 from rwkv.model import RWKV
 model_path = hf_hub_download(repo_id="BlinkDL/rwkv-4-pile-14b", filename="RWKV-4-Pile-14B-20230228-ctx4096-test663.pth")
-model = RWKV(model=model_path, strategy='cuda fp16 *32 -> cpu fp32')
-# model_path = hf_hub_download(repo_id="BlinkDL/rwkv-4-pile-169m", filename="RWKV-4-Pile-169M-20220807-8023.pth")
-# model = RWKV(model=model_path, strategy='cuda fp16')
+model = RWKV(model=model_path, strategy='cuda fp16i8 *10 -> cuda fp16')
 from rwkv.utils import PIPELINE, PIPELINE_ARGS
 pipeline = PIPELINE(model, "20B_tokenizer.json")