Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
|
@@ -63,11 +63,16 @@ SPACE_REPO_ID = os.environ.get("SPACE_REPO_ID", "")
|
|
| 63 |
# Generation / toggles
|
| 64 |
ALLOW_WIKIPEDIA = False
|
| 65 |
DEBUG = True
|
| 66 |
-
MAX_NEW_TOKENS_GROUNDED =
|
| 67 |
-
MAX_NEW_TOKENS_FALLBACK =
|
| 68 |
-
|
| 69 |
MIN_USEFUL_CHARS = 260
|
| 70 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 71 |
def dlog(tag, msg):
|
| 72 |
if DEBUG: print(f"[{tag}] {msg}")
|
| 73 |
|
|
@@ -267,8 +272,35 @@ GEN_ARGS_FALLBACK = dict(
|
|
| 267 |
|
| 268 |
def _generate(inputs, grounded: bool):
|
| 269 |
args = GEN_ARGS_GROUNDED if grounded else GEN_ARGS_FALLBACK
|
|
|
|
| 270 |
with torch.inference_mode():
|
| 271 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 272 |
|
| 273 |
# ================== UTILITIES ==================
|
| 274 |
_SANITIZE = re.compile(r"```.*?```|<\s*script[^>]*>.*?<\s*/\s*script\s*>", re.DOTALL|re.IGNORECASE)
|
|
@@ -986,45 +1018,43 @@ with gr.Blocks(theme="soft") as demo:
|
|
| 986 |
submit_fb = gr.Button("Submit feedback")
|
| 987 |
fb_status = gr.Markdown("")
|
| 988 |
|
| 989 |
-
|
| 990 |
-
|
| 991 |
-
|
| 992 |
-
|
| 993 |
-
|
| 994 |
-
|
| 995 |
|
| 996 |
-
|
| 997 |
-
predict,
|
| 998 |
inputs=[user_in, chat, state],
|
| 999 |
outputs=[chat, user_in, feedback_grp, rating, comment, state],
|
| 1000 |
-
concurrency_limit=
|
| 1001 |
-
|
| 1002 |
-
|
| 1003 |
|
| 1004 |
-
|
| 1005 |
-
predict,
|
| 1006 |
inputs=[user_in, chat, state],
|
| 1007 |
outputs=[chat, user_in, feedback_grp, rating, comment, state],
|
| 1008 |
-
concurrency_limit=
|
| 1009 |
-
|
| 1010 |
-
|
| 1011 |
-
clear_btn.click(
|
| 1012 |
-
lambda: ([], "", gr.update(visible=False), None, "", init_session()),
|
| 1013 |
-
inputs=None,
|
| 1014 |
-
outputs=[chat, user_in, feedback_grp, rating, comment, state],
|
| 1015 |
-
concurrency_limit=4,
|
| 1016 |
-
)
|
| 1017 |
|
| 1018 |
-
|
| 1019 |
-
|
| 1020 |
-
|
| 1021 |
-
|
| 1022 |
-
|
| 1023 |
-
|
| 1024 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1025 |
|
|
|
|
|
|
|
| 1026 |
|
| 1027 |
-
demo.queue(max_size=64)
|
| 1028 |
-
demo.launch(max_threads=int(os.environ.get("MAX_THREADS", "32")))
|
| 1029 |
|
| 1030 |
|
|
|
|
| 63 |
# Generation / toggles
|
| 64 |
ALLOW_WIKIPEDIA = False
|
| 65 |
DEBUG = True
|
| 66 |
+
MAX_NEW_TOKENS_GROUNDED = 512
|
| 67 |
+
MAX_NEW_TOKENS_FALLBACK = 256
|
|
|
|
| 68 |
MIN_USEFUL_CHARS = 260
|
| 69 |
|
| 70 |
+
# Auto-continue if we hit the cap without EOS
|
| 71 |
+
AUTO_CONTINUE = True
|
| 72 |
+
AUTO_CONT_MAX_STEPS = 2 # continue up to 2 extra chunks
|
| 73 |
+
AUTO_CONT_NEW_TOKENS = 256 # tokens per continuation step
|
| 74 |
+
|
| 75 |
+
|
| 76 |
def dlog(tag, msg):
|
| 77 |
if DEBUG: print(f"[{tag}] {msg}")
|
| 78 |
|
|
|
|
| 272 |
|
| 273 |
def _generate(inputs, grounded: bool):
|
| 274 |
args = GEN_ARGS_GROUNDED if grounded else GEN_ARGS_FALLBACK
|
| 275 |
+
in_len = inputs["input_ids"].shape[-1]
|
| 276 |
with torch.inference_mode():
|
| 277 |
+
out = model_lm.generate(**inputs, **args)
|
| 278 |
+
|
| 279 |
+
if not AUTO_CONTINUE:
|
| 280 |
+
return out
|
| 281 |
+
|
| 282 |
+
steps = 0
|
| 283 |
+
while steps < AUTO_CONT_MAX_STEPS:
|
| 284 |
+
seq = out[0]
|
| 285 |
+
ended_with_eos = (seq[-1].item() == tokenizer_lm.eos_token_id)
|
| 286 |
+
hit_cap = (seq.shape[0] - in_len) >= args["max_new_tokens"]
|
| 287 |
+
if ended_with_eos or not hit_cap:
|
| 288 |
+
break
|
| 289 |
+
|
| 290 |
+
# continue generation from the current sequence
|
| 291 |
+
cont_inputs = {
|
| 292 |
+
"input_ids": seq.unsqueeze(0),
|
| 293 |
+
"attention_mask": torch.ones_like(seq).unsqueeze(0),
|
| 294 |
+
}
|
| 295 |
+
cont_inputs = {k: v.to(device) for k, v in cont_inputs.items()}
|
| 296 |
+
cont_args = dict(args)
|
| 297 |
+
cont_args["max_new_tokens"] = AUTO_CONT_NEW_TOKENS
|
| 298 |
+
|
| 299 |
+
out = model_lm.generate(**cont_inputs, **cont_args)
|
| 300 |
+
steps += 1
|
| 301 |
+
|
| 302 |
+
return out
|
| 303 |
+
|
| 304 |
|
| 305 |
# ================== UTILITIES ==================
|
| 306 |
_SANITIZE = re.compile(r"```.*?```|<\s*script[^>]*>.*?<\s*/\s*script\s*>", re.DOTALL|re.IGNORECASE)
|
|
|
|
| 1018 |
submit_fb = gr.Button("Submit feedback")
|
| 1019 |
fb_status = gr.Markdown("")
|
| 1020 |
|
| 1021 |
+
# Wiring
|
| 1022 |
+
enter_btn.click(
|
| 1023 |
+
fn=enter_app,
|
| 1024 |
+
inputs=[first_tb, last_tb, state],
|
| 1025 |
+
outputs=[gate, app, state, gate_msg],
|
| 1026 |
+
)
|
| 1027 |
|
| 1028 |
+
send_btn.click(
|
| 1029 |
+
fn=predict,
|
| 1030 |
inputs=[user_in, chat, state],
|
| 1031 |
outputs=[chat, user_in, feedback_grp, rating, comment, state],
|
| 1032 |
+
concurrency_limit=1, # serialize LLM calls
|
| 1033 |
+
)
|
|
|
|
| 1034 |
|
| 1035 |
+
user_in.submit(
|
| 1036 |
+
fn=predict,
|
| 1037 |
inputs=[user_in, chat, state],
|
| 1038 |
outputs=[chat, user_in, feedback_grp, rating, comment, state],
|
| 1039 |
+
concurrency_limit=1, # serialize LLM calls
|
| 1040 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1041 |
|
| 1042 |
+
clear_btn.click(
|
| 1043 |
+
fn=lambda: ([], "", gr.update(visible=False), None, "", init_session()),
|
| 1044 |
+
inputs=None,
|
| 1045 |
+
outputs=[chat, user_in, feedback_grp, rating, comment, state],
|
| 1046 |
+
concurrency_limit=4,
|
| 1047 |
+
)
|
| 1048 |
|
| 1049 |
+
submit_fb.click(
|
| 1050 |
+
fn=save_feedback,
|
| 1051 |
+
inputs=[rating, comment, state],
|
| 1052 |
+
outputs=[fb_status, feedback_grp],
|
| 1053 |
+
concurrency_limit=4,
|
| 1054 |
+
)
|
| 1055 |
|
| 1056 |
+
# Queue (true concurrency = 1 to prevent OOM/restarts)
|
| 1057 |
+
demo.queue(concurrency_count=1, max_size=64)
|
| 1058 |
|
|
|
|
|
|
|
| 1059 |
|
| 1060 |
|