vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

updated lite, fixed some encoding issues

Browse source
This commit is contained in:
Concedo 2023-05-21 17:29:00 +08:00
parent 75e4548821
commit 24127ebf98
3 changed files with 28 additions and 16 deletions
Download patch file Download diff file Expand all files Collapse all files

14
klite.embd
View file

File diff suppressed because one or more lines are too long

14
koboldcpp.py
View file

@ -191,6 +191,14 @@ def generate(prompt,max_length=20, max_context_length=512,temperature=0.8,top_k=
return ret.text.decode("UTF-8","ignore")
return ""
def utfprint(str):
try:
print(str)
except UnicodeEncodeError:
# Replace or omit the problematic character
utf_string = str.encode('ascii', 'ignore').decode('ascii')
print(utf_string)
#################################################################
### A hacky simple HTTP server simulating a kobold api by Concedo
### we are intentionally NOT using flask, because we want MINIMAL dependencies
@ -301,7 +309,7 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
self.send_response(503)
self.end_headers()
return
print("\nInput: " + json.dumps(genparams))
utfprint("\nInput: " + json.dumps(genparams))
modelbusy = True
if kai_api_flag:
@ -327,7 +335,7 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
seed=-1,
stop_sequence=genparams.get('stop_sequence', [])
)
print("\nOutput: " + recvtxt)
utfprint("\nOutput: " + recvtxt)
res = {"results": [{"text": recvtxt}]}
else:
recvtxt = generate(
@ -343,7 +351,7 @@ class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):
seed=-1,
stop_sequence=genparams.get('stop_sequence', [])
)
print("\nOutput: " + recvtxt)
utfprint("\nOutput: " + recvtxt)
res = {"data": {"seqs":[recvtxt]}}
try:

16
otherarch/llama_v2.cpp
View file

@ -1071,17 +1071,17 @@ static void llama_v2_model_load_internal(
for (int i = 0; i < n_gpu; ++i) {
const auto & layer = model.layers[i];
ggml_v2_cuda_transform_tensor(layer.wq); vram_total += ggml_v2_nbytes(layer.wq);
ggml_v2_cuda_transform_tensor(layer.wk); vram_total += ggml_v2_nbytes(layer.wk);
ggml_v2_cuda_transform_tensor(layer.wv); vram_total += ggml_v2_nbytes(layer.wv);
ggml_v2_cuda_transform_tensor(layer.wo); vram_total += ggml_v2_nbytes(layer.wo);
ggml_v2_cuda_transform_tensor(layer.w1); vram_total += ggml_v2_nbytes(layer.w1);
ggml_v2_cuda_transform_tensor(layer.w2); vram_total += ggml_v2_nbytes(layer.w2);
ggml_v2_cuda_transform_tensor(layer.w3); vram_total += ggml_v2_nbytes(layer.w3);
ggml_cuda_transform_tensor(layer.wq); vram_total += ggml_v2_nbytes(layer.wq);
ggml_cuda_transform_tensor(layer.wk); vram_total += ggml_v2_nbytes(layer.wk);
ggml_cuda_transform_tensor(layer.wv); vram_total += ggml_v2_nbytes(layer.wv);
ggml_cuda_transform_tensor(layer.wo); vram_total += ggml_v2_nbytes(layer.wo);
ggml_cuda_transform_tensor(layer.w1); vram_total += ggml_v2_nbytes(layer.w1);
ggml_cuda_transform_tensor(layer.w2); vram_total += ggml_v2_nbytes(layer.w2);
ggml_cuda_transform_tensor(layer.w3); vram_total += ggml_v2_nbytes(layer.w3);
}
if (n_gpu_layers > (int) hparams.n_layer) {
fprintf(stderr, "%s: [cublas] offloading output layer to GPU\n", __func__);
ggml_v2_cuda_transform_tensor(model.output); vram_total += ggml_v2_nbytes(model.output);
ggml_cuda_transform_tensor(model.output); vram_total += ggml_v2_nbytes(model.output);
}
fprintf(stderr, "%s: [cublas] total VRAM used: %zu MB\n", __func__, vram_total / 1024 / 1024);