Spaces:
Running
on
T4
Running
on
T4
Update app.py
Browse files
app.py
CHANGED
|
@@ -77,12 +77,22 @@ def evaluate(
|
|
| 77 |
state = None
|
| 78 |
for i in range(int(token_count)):
|
| 79 |
|
| 80 |
-
|
| 81 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
for n in occurrence:
|
| 83 |
-
|
| 84 |
|
| 85 |
-
token = pipeline.sample_logits(
|
| 86 |
if token in args.token_stop:
|
| 87 |
break
|
| 88 |
all_tokens += [token]
|
|
@@ -91,8 +101,8 @@ def evaluate(
|
|
| 91 |
|
| 92 |
ttt = pipeline.decode([token])
|
| 93 |
www = 1
|
| 94 |
-
if ttt in ' \t0123456789':
|
| 95 |
-
|
| 96 |
#elif ttt in '\r\n,.;?!"\':+-*/=#@$%^&_`~|<>\\()[]{},。;“”:?!()【】':
|
| 97 |
# www = 0.5
|
| 98 |
if token not in occurrence:
|
|
@@ -138,7 +148,7 @@ with gr.Blocks(title=title, theme=gr.themes.Base()) as demo:
|
|
| 138 |
gr.HTML(f"<div style=\"text-align: center;\">\n<h1>{title}</h1>\n</div>")
|
| 139 |
|
| 140 |
with gr.Tab("=== Base Model (Raw Generation) ==="):
|
| 141 |
-
gr.Markdown(f'This is [RWKV7 G-series](https://huggingface.co/BlinkDL/rwkv7-g1) 2.9B reasoning base LM - an attention-free pure RNN [RWKV-LM](https://github.com/BlinkDL/RWKV-LM). Try topp0 penalty0 for math/code/translation. Supports 100+ world languages and code. Check [
|
| 142 |
with gr.Row():
|
| 143 |
with gr.Column():
|
| 144 |
prompt = gr.Textbox(lines=6, label="Prompt", value="User: simulate SpaceX mars landing using python\n\nAssistant: <think")
|
|
|
|
| 77 |
state = None
|
| 78 |
for i in range(int(token_count)):
|
| 79 |
|
| 80 |
+
if i == 0:
|
| 81 |
+
input_ids = pipeline.encode(ctx)[-ctx_limit:]
|
| 82 |
+
out, state = model.forward(input_ids, state)
|
| 83 |
+
for j in range(len(state)):
|
| 84 |
+
static_state_in[j].copy_(state[j])
|
| 85 |
+
static_output.copy_(out)
|
| 86 |
+
else:
|
| 87 |
+
static_input.copy_(model.z['emb.weight'][token])
|
| 88 |
+
graph.replay()
|
| 89 |
+
for j in range(len(state)):
|
| 90 |
+
static_state_in[j].copy_(static_state_out[j])
|
| 91 |
+
|
| 92 |
for n in occurrence:
|
| 93 |
+
static_output[n] -= (args.alpha_presence + occurrence[n] * args.alpha_frequency)
|
| 94 |
|
| 95 |
+
token = pipeline.sample_logits(static_output, temperature=args.temperature, top_p=args.top_p)
|
| 96 |
if token in args.token_stop:
|
| 97 |
break
|
| 98 |
all_tokens += [token]
|
|
|
|
| 101 |
|
| 102 |
ttt = pipeline.decode([token])
|
| 103 |
www = 1
|
| 104 |
+
#if ttt in ' \t0123456789':
|
| 105 |
+
# www = 0
|
| 106 |
#elif ttt in '\r\n,.;?!"\':+-*/=#@$%^&_`~|<>\\()[]{},。;“”:?!()【】':
|
| 107 |
# www = 0.5
|
| 108 |
if token not in occurrence:
|
|
|
|
| 148 |
gr.HTML(f"<div style=\"text-align: center;\">\n<h1>{title}</h1>\n</div>")
|
| 149 |
|
| 150 |
with gr.Tab("=== Base Model (Raw Generation) ==="):
|
| 151 |
+
gr.Markdown(f'This is [RWKV7 G-series](https://huggingface.co/BlinkDL/rwkv7-g1) 2.9B reasoning base LM - an attention-free pure RNN [RWKV-LM](https://github.com/BlinkDL/RWKV-LM). Try topp0 penalty0 for math/code/translation. Supports 100+ world languages and code. Check [600+ Github RWKV projects](https://github.com/search?o=desc&p=1&q=rwkv&s=updated&type=Repositories). *** Can try examples (bottom of page) *** (can edit them). Demo limited to ctxlen {ctx_limit}.')
|
| 152 |
with gr.Row():
|
| 153 |
with gr.Column():
|
| 154 |
prompt = gr.Textbox(lines=6, label="Prompt", value="User: simulate SpaceX mars landing using python\n\nAssistant: <think")
|