update
Browse files
app.py
CHANGED
@@ -3,6 +3,7 @@ import math
|
|
3 |
import os
|
4 |
import time
|
5 |
from threading import Thread
|
|
|
6 |
|
7 |
import gradio as gr
|
8 |
import spaces
|
@@ -18,7 +19,6 @@ from langchain_docling.loader import ExportType
|
|
18 |
from langchain_text_splitters import RecursiveCharacterTextSplitter
|
19 |
from transformers import AutoModelForCausalLM, AutoTokenizer, DynamicCache, TextIteratorStreamer
|
20 |
from transformers.models.llama.modeling_llama import rotate_half
|
21 |
-
import uuid
|
22 |
|
23 |
from utils import (
|
24 |
calculate_tokens_suggest_compression_ratio,
|
@@ -26,8 +26,6 @@ from utils import (
|
|
26 |
update_retrieval_context,
|
27 |
)
|
28 |
|
29 |
-
|
30 |
-
|
31 |
# Initialize the model and tokenizer.
|
32 |
api_token = os.getenv("HUGGING_FACE_HUB_TOKEN")
|
33 |
model_name = "meta-llama/Llama-3.1-8B-Instruct"
|
@@ -37,12 +35,11 @@ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
37 |
model = model.eval()
|
38 |
model.to(device)
|
39 |
embedding_model = HuggingFaceBgeEmbeddings(
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
|
47 |
# Create a chat template and split into prefix and suffix.
|
48 |
content_system = ""
|
@@ -121,7 +118,6 @@ class FinchCache(DynamicCache):
|
|
121 |
self._seen_tokens = new_length
|
122 |
|
123 |
def save(self, path: str):
|
124 |
-
"""Save the cache to disk, moving tensors to CPU."""
|
125 |
try:
|
126 |
os.makedirs(os.path.dirname(path), exist_ok=True)
|
127 |
torch.save(
|
@@ -133,7 +129,6 @@ class FinchCache(DynamicCache):
|
|
133 |
|
134 |
@classmethod
|
135 |
def load(cls, path: str, device: str = "cpu") -> "FinchCache":
|
136 |
-
"""Load the cache from disk and move tensors to the specified device."""
|
137 |
data = torch.load(path, map_location=device)
|
138 |
cache = cls()
|
139 |
cache.key_cache = [k.to(device) for k in data["key_cache"]]
|
@@ -141,8 +136,6 @@ class FinchCache(DynamicCache):
|
|
141 |
cache._seen_tokens = cache.value_cache[0].size(2) if cache.value_cache else 0
|
142 |
return cache
|
143 |
|
144 |
-
|
145 |
-
|
146 |
def convert_to_markdown(file_objs, url, do_ocr, do_table_structure):
|
147 |
file_path = file_objs if file_objs is not None else url
|
148 |
pipeline_options = PdfPipelineOptions()
|
@@ -154,12 +147,8 @@ def convert_to_markdown(file_objs, url, do_ocr, do_table_structure):
|
|
154 |
)
|
155 |
doc_converter = DocumentConverter(
|
156 |
allowed_formats=[InputFormat.PDF],
|
157 |
-
format_options={
|
158 |
-
InputFormat.PDF: pdf_format_options
|
159 |
-
}
|
160 |
)
|
161 |
-
|
162 |
-
# Pass the custom converter to the DoclingLoader.
|
163 |
loader = DoclingLoader(
|
164 |
file_path=file_path,
|
165 |
export_type=ExportType.MARKDOWN,
|
@@ -168,39 +157,51 @@ def convert_to_markdown(file_objs, url, do_ocr, do_table_structure):
|
|
168 |
docs = loader.load()
|
169 |
return docs[0].page_content
|
170 |
|
171 |
-
|
172 |
def create_rag_index(collection_name, text_no_prefix):
|
173 |
-
"""Loads the PDF, splits its text, and builds a vectorstore for naive RAG."""
|
174 |
text_splitter = RecursiveCharacterTextSplitter.from_huggingface_tokenizer(
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
# Concatenate pages and create Document objects.
|
183 |
docs = [Document(page_content=x) for x in text_splitter.split_text(text_no_prefix)]
|
184 |
vectorstore = Chroma.from_documents(collection_name=collection_name, persist_directory="./chroma_db", documents=docs, embedding=embedding_model)
|
185 |
return vectorstore
|
186 |
|
187 |
-
|
188 |
@spaces.GPU
|
189 |
def auto_convert(file_objs, url, do_ocr, do_table_structure):
|
|
|
|
|
190 |
if file_objs is None and (url is None or url.strip() == ""):
|
191 |
return (
|
192 |
gr.update(value=""),
|
193 |
"Number of tokens before compression: ",
|
194 |
-
gr.update(),
|
195 |
"Number of tokens after compression: ",
|
196 |
0,
|
197 |
-
gr.update(interactive=False),
|
198 |
False,
|
199 |
-
{}
|
|
|
200 |
)
|
201 |
-
# Convert the document to markdown.
|
202 |
print("Converting to markdown")
|
203 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
204 |
print("Done")
|
205 |
combined_text = prefix + markdown
|
206 |
print("Suggestioning Compression ratio")
|
@@ -213,8 +214,6 @@ def auto_convert(file_objs, url, do_ocr, do_table_structure):
|
|
213 |
token_count_str = f"Number of tokens before compression: {token_count}"
|
214 |
retrieval_str = f"Number of tokens after compression: {retrieval_tokens}"
|
215 |
slider_update = gr.update(value=default_ratio, minimum=min_ratio, maximum=max_ratio, step=1)
|
216 |
-
|
217 |
-
# Create the RAG index immediately.
|
218 |
if combined_text.startswith(prefix):
|
219 |
rag_text = combined_text[len(prefix):]
|
220 |
else:
|
@@ -223,18 +222,17 @@ def auto_convert(file_objs, url, do_ocr, do_table_structure):
|
|
223 |
rag_index = create_rag_index(collection_name, rag_text)
|
224 |
state = {"rag_index": collection_name}
|
225 |
print("Done")
|
226 |
-
|
227 |
return (
|
228 |
-
combined_text,
|
229 |
-
token_count_str,
|
230 |
-
slider_update,
|
231 |
-
retrieval_str,
|
232 |
-
token_count,
|
233 |
-
gr.update(interactive=True),
|
234 |
False,
|
235 |
-
state
|
|
|
236 |
)
|
237 |
-
|
238 |
|
239 |
def get_compressed_kv_cache(sink_tokens, step_size, target_token_size, context_ids, context_attention_mask, question_ids, question_attention_mask):
|
240 |
device = model.device
|
@@ -250,32 +248,18 @@ def get_compressed_kv_cache(sink_tokens, step_size, target_token_size, context_i
|
|
250 |
max_context_tokens_allowed = model.config.max_position_embeddings - question_len
|
251 |
if total_len > max_context_tokens_allowed:
|
252 |
num_chunks = max(step_size, math.ceil(total_len / max_context_tokens_allowed))
|
253 |
-
|
254 |
if total_len <= sink_tokens or num_chunks == 1:
|
255 |
-
# If the context is too short or only one chunk is desired, use the entire context.
|
256 |
context_ids_list = [context_ids]
|
257 |
context_attention_mask_list = [context_attention_mask]
|
258 |
else:
|
259 |
-
# Calculate how many tokens remain after the sink tokens.
|
260 |
remainder_len = total_len - sink_tokens
|
261 |
-
|
262 |
-
# Compute the base tokens per chunk and any leftover.
|
263 |
base = remainder_len // num_chunks
|
264 |
leftover = remainder_len % num_chunks
|
265 |
-
|
266 |
-
# Build a list of chunk sizes.
|
267 |
-
# First chunk gets the sink tokens plus base tokens.
|
268 |
chunk_sizes = [sink_tokens + base]
|
269 |
-
|
270 |
-
# Chunks 2 to num_chunks-1 get base tokens each.
|
271 |
for _ in range(num_chunks - 2):
|
272 |
chunk_sizes.append(base)
|
273 |
-
|
274 |
-
# The last chunk gets the remaining tokens (base + leftover).
|
275 |
if num_chunks > 1:
|
276 |
chunk_sizes.append(base + leftover)
|
277 |
-
|
278 |
-
# Now slice the context using the calculated sizes.
|
279 |
context_ids_list = []
|
280 |
context_attention_mask_list = []
|
281 |
offset = 0
|
@@ -284,33 +268,23 @@ def get_compressed_kv_cache(sink_tokens, step_size, target_token_size, context_i
|
|
284 |
context_ids_list.append(context_ids[:, offset:end])
|
285 |
context_attention_mask_list.append(context_attention_mask[:, offset:end])
|
286 |
offset = end
|
287 |
-
|
288 |
-
# (Optional) Continue with the rest of your processing…
|
289 |
len_rest = max(total_len - sink_tokens, 1)
|
290 |
compression_factor = len_rest // target_token_size
|
291 |
if compression_factor < 1:
|
292 |
compression_factor = 1
|
293 |
-
|
294 |
tokenized_doc_chunks = []
|
295 |
for ids_chunk, mask_chunk in zip(context_ids_list, context_attention_mask_list):
|
296 |
tokenized_doc_chunks.append({"input_ids": ids_chunk, "attention_mask": mask_chunk})
|
297 |
-
|
298 |
print("Number of chunks: ", len(tokenized_doc_chunks))
|
299 |
-
|
300 |
rotary_emb = model.model.rotary_emb.to(device)
|
301 |
inv_freq = rotary_emb.inv_freq
|
302 |
batch_size = question_ids.size(0)
|
303 |
ones_mask = torch.ones(batch_size, 1, dtype=question_attention_mask.dtype, device=device)
|
304 |
-
|
305 |
cache = FinchCache()
|
306 |
past_cache_len = 0
|
307 |
past_attention_mask = torch.zeros(batch_size, 0, dtype=question_attention_mask.dtype, device=device)
|
308 |
num_chunks = len(tokenized_doc_chunks)
|
309 |
-
|
310 |
-
# Prepare a shared dictionary for hook outputs.
|
311 |
query_context_matrices = {}
|
312 |
-
|
313 |
-
# Define a hook function that uses a per-chunk offset stored on self.
|
314 |
def query_hook_fn(module, input, output):
|
315 |
layer_idx = getattr(module, "layer_idx", None)
|
316 |
if layer_idx is not None:
|
@@ -323,26 +297,17 @@ def get_compressed_kv_cache(sink_tokens, step_size, target_token_size, context_i
|
|
323 |
.transpose(1, 2)
|
324 |
.contiguous()
|
325 |
)
|
326 |
-
# Use self._current_chunk_offset to select only the new tokens.
|
327 |
query_context_matrices[layer_idx] = query_states[:, :, _current_chunk_offset:, :].clone()
|
328 |
-
|
329 |
-
# Pre-register hooks for all layers only once.
|
330 |
hooks = []
|
331 |
for i, layer in enumerate(model.model.layers):
|
332 |
-
layer.self_attn.q_proj.layer_idx = i
|
333 |
layer.self_attn.q_proj.num_query_heads = layer.self_attn.config.num_attention_heads
|
334 |
hook = layer.self_attn.q_proj.register_forward_hook(query_hook_fn)
|
335 |
hooks.append(hook)
|
336 |
-
|
337 |
-
# Process each document chunk sequentially.
|
338 |
for j, tokenized_doc_chunk in enumerate(tokenized_doc_chunks):
|
339 |
current_seq_length = tokenized_doc_chunk["input_ids"].size(1)
|
340 |
-
# Save the offset in an attribute the hook can access.
|
341 |
_current_chunk_offset = current_seq_length
|
342 |
-
# Clear the dictionary from any previous chunk.
|
343 |
query_context_matrices.clear()
|
344 |
-
|
345 |
-
# These chunks are already on the device.
|
346 |
chunk_input_ids = tokenized_doc_chunk["input_ids"].contiguous()
|
347 |
chunk_attention_mask = tokenized_doc_chunk["attention_mask"].contiguous()
|
348 |
segment_attention_mask = torch.cat(
|
@@ -350,7 +315,6 @@ def get_compressed_kv_cache(sink_tokens, step_size, target_token_size, context_i
|
|
350 |
).contiguous()
|
351 |
current_input_ids = torch.cat([chunk_input_ids, question_ids], dim=-1).contiguous()
|
352 |
current_attention_mask = torch.cat([segment_attention_mask, question_attention_mask], dim=-1).contiguous()
|
353 |
-
|
354 |
past_seen_tokens = cache.get_seq_length() if cache is not None else 0
|
355 |
cache_position = torch.arange(
|
356 |
past_seen_tokens + chunk_input_ids.shape[1],
|
@@ -366,7 +330,6 @@ def get_compressed_kv_cache(sink_tokens, step_size, target_token_size, context_i
|
|
366 |
cache_position=cache_position,
|
367 |
batch_size=current_input_ids.size(0),
|
368 |
).contiguous()
|
369 |
-
|
370 |
with torch.no_grad():
|
371 |
outputs = model.model(
|
372 |
input_ids=current_input_ids,
|
@@ -374,9 +337,7 @@ def get_compressed_kv_cache(sink_tokens, step_size, target_token_size, context_i
|
|
374 |
past_key_values=cache,
|
375 |
)
|
376 |
cache = outputs.past_key_values
|
377 |
-
|
378 |
len_question = question_ids.size(1)
|
379 |
-
# Now, for each transformer layer, update the cache using the query/key attention.
|
380 |
for layer_idx in range(len(model.model.layers)):
|
381 |
key_matrix = cache.key_cache[layer_idx]
|
382 |
query_matrix = query_context_matrices[layer_idx]
|
@@ -392,13 +353,11 @@ def get_compressed_kv_cache(sink_tokens, step_size, target_token_size, context_i
|
|
392 |
query_matrix = (query_matrix * cos) + (rotate_half(query_matrix) * sin)
|
393 |
num_repeats = model.config.num_attention_heads // model.config.num_key_value_heads
|
394 |
key_matrix = repeat_kv(key_matrix, num_repeats)
|
395 |
-
|
396 |
scaling = math.sqrt(model.config.head_dim)
|
397 |
attention_matrix = torch.matmul(query_matrix, key_matrix.transpose(2, 3)) / scaling
|
398 |
causal_mask_sliced = causal_mask[:, :, :, : key_matrix.shape[-2]]
|
399 |
attention_matrix = attention_matrix + causal_mask_sliced
|
400 |
attention_matrix = torch.nn.functional.softmax(attention_matrix, dim=-1, dtype=torch.float32).to(query_matrix.dtype)
|
401 |
-
# Normalization
|
402 |
tol = 1e-8
|
403 |
binary_mask = (torch.abs(causal_mask_sliced.to(torch.float32)) < tol).to(torch.float32)
|
404 |
non_zero_counts = binary_mask.sum(dim=3, keepdim=True)
|
@@ -423,30 +382,21 @@ def get_compressed_kv_cache(sink_tokens, step_size, target_token_size, context_i
|
|
423 |
to_keep_new = int(current_seq_length // compression_factor)
|
424 |
k = min(past_cache_len + to_keep_new, full_context_size)
|
425 |
else:
|
426 |
-
desired_final = sink_tokens + target_token_size + len_question
|
427 |
k = desired_final if full_context_size >= desired_final else full_context_size
|
428 |
k = max(k, sink_tokens)
|
429 |
selected_indices = torch.topk(attention_matrix, k, dim=-1).indices
|
430 |
selected_indices, _ = torch.sort(selected_indices, dim=-1)
|
431 |
cache.compress_cache(layer_idx, selected_indices, inv_freq)
|
432 |
-
|
433 |
past_cache_len = cache._seen_tokens
|
434 |
past_attention_mask = torch.ones(1, past_cache_len, device=device)
|
435 |
-
|
436 |
-
# Remove the hooks once after all chunks are processed.
|
437 |
for hook in hooks:
|
438 |
hook.remove()
|
439 |
-
|
440 |
return cache
|
441 |
|
442 |
-
|
443 |
def run_naive_rag_query(collection_name, query, rag_token_size, prefix, task, few_shot_examples):
|
444 |
-
"""
|
445 |
-
For naive RAG, retrieves top-k chunks (k based on target token size)
|
446 |
-
and generates an answer using those chunks.
|
447 |
-
"""
|
448 |
k = max(1, rag_token_size // 256)
|
449 |
-
vectorstore = Chroma(persist_directory="./chroma_db",
|
450 |
retriever = vectorstore.as_retriever(search_type="similarity", search_kwargs={"k": k})
|
451 |
retrieved_docs = retriever.invoke(query)
|
452 |
for doc in retrieved_docs:
|
@@ -454,17 +404,11 @@ def run_naive_rag_query(collection_name, query, rag_token_size, prefix, task, fe
|
|
454 |
print(doc.page_content)
|
455 |
print("=================")
|
456 |
formatted_context = "\n\n".join([doc.page_content for doc in retrieved_docs])
|
457 |
-
|
458 |
rag_context = prefix + "Retrieved context: \n" + formatted_context + task + few_shot_examples
|
459 |
-
|
460 |
return rag_context
|
461 |
|
462 |
-
|
463 |
@spaces.GPU
|
464 |
def prepare_compression_and_rag(combined_text, retrieval_slider_value, global_local_value, task_description, few_shot, state):
|
465 |
-
"""
|
466 |
-
Prepares the compressed KV cache. Uses the precomputed rag_index from state.
|
467 |
-
"""
|
468 |
percentage = int(global_local_value.replace('%', ''))
|
469 |
question_text = task_description + "\n" + few_shot
|
470 |
context_encoding = tokenizer(combined_text, return_tensors="pt").to(device)
|
@@ -474,15 +418,23 @@ def prepare_compression_and_rag(combined_text, retrieval_slider_value, global_lo
|
|
474 |
question_ids = question_encoding["input_ids"]
|
475 |
question_attention_mask = question_encoding["attention_mask"]
|
476 |
retrieval_context_length = int(context_ids.size(1) / retrieval_slider_value)
|
477 |
-
|
|
|
|
|
|
|
478 |
if percentage > 0:
|
479 |
target_token_size = int(retrieval_context_length * (percentage / 100))
|
480 |
print("Target token size for compression: ", target_token_size)
|
481 |
step_size = 2
|
482 |
start_time_prefill = time.perf_counter()
|
483 |
-
|
484 |
-
|
485 |
-
|
|
|
|
|
|
|
|
|
|
|
486 |
compressed_length = past_key_values.get_seq_length()
|
487 |
print("Context size after compression: ", compressed_length)
|
488 |
print("Compression rate: ", context_ids.size(1) / compressed_length)
|
@@ -491,15 +443,11 @@ def prepare_compression_and_rag(combined_text, retrieval_slider_value, global_lo
|
|
491 |
target_token_size = 0
|
492 |
past_key_values = FinchCache()
|
493 |
compressed_length = past_key_values.get_seq_length()
|
494 |
-
|
495 |
-
cache_name = "default_cache_" + uuid.uuid4().hex[:6]
|
496 |
cache_name = "default_cache_" + uuid.uuid4().hex[:6] + ".pt"
|
497 |
save_dir = "./cache_dir"
|
498 |
os.makedirs(save_dir, exist_ok=True)
|
499 |
save_path = os.path.join(save_dir, cache_name)
|
500 |
past_key_values.save(save_path)
|
501 |
-
|
502 |
-
# Use the precomputed rag_index from state.
|
503 |
collection_name = state.get("rag_index", None)
|
504 |
if collection_name is None:
|
505 |
print("Collection name not found creating a new one.")
|
@@ -509,7 +457,6 @@ def prepare_compression_and_rag(combined_text, retrieval_slider_value, global_lo
|
|
509 |
rag_text = combined_text
|
510 |
collection_name = "default_collection_" + uuid.uuid4().hex[:6]
|
511 |
rag_index = create_rag_index(collection_name, rag_text)
|
512 |
-
|
513 |
state.update({
|
514 |
"compressed_cache": save_path,
|
515 |
"compressed_length": compressed_length,
|
@@ -520,32 +467,28 @@ def prepare_compression_and_rag(combined_text, retrieval_slider_value, global_lo
|
|
520 |
"task_description": task_description,
|
521 |
"few_shot": few_shot,
|
522 |
"retrieval_slider": retrieval_context_length,
|
523 |
-
"prefill_time": time.perf_counter() - start_time_prefill
|
|
|
|
|
|
|
524 |
})
|
525 |
return state, True
|
526 |
|
527 |
-
|
528 |
@spaces.GPU
|
529 |
def chat_response_stream(message: str, history: list, state: dict):
|
530 |
-
|
531 |
-
|
532 |
-
|
533 |
-
|
534 |
user_message = message
|
535 |
save_path = state["compressed_cache"]
|
536 |
past_key_values = FinchCache.load(save_path, device=model.device)
|
537 |
-
try:
|
538 |
-
os.remove(save_path)
|
539 |
-
except Exception as e:
|
540 |
-
print(f"Error removing cache file: {e}")
|
541 |
compressed_length = past_key_values.get_seq_length()
|
542 |
collection_name = state["rag_index"]
|
543 |
retrieval_slider_value = state["retrieval_slider"]
|
544 |
percentage = state["global_local"]
|
545 |
-
|
546 |
rag_retrieval_size = int(retrieval_slider_value * (1.0 - (percentage / 100)))
|
547 |
print("RAG retrieval size: ", rag_retrieval_size)
|
548 |
-
|
549 |
if percentage == 0:
|
550 |
rag_prefix = prefix
|
551 |
rag_task = state["task_description"]
|
@@ -565,7 +508,6 @@ def chat_response_stream(message: str, history: list, state: dict):
|
|
565 |
eos_block = torch.full((1, compressed_length), tokenizer.eos_token_id, device=device, dtype=torch.long)
|
566 |
new_input_ids = torch.cat([eos_block, tokenized_new_input["input_ids"]], dim=-1)
|
567 |
new_attention_mask = torch.cat([torch.ones((1, compressed_length), device=device), tokenized_new_input["attention_mask"]], dim=-1)
|
568 |
-
|
569 |
print("New input is: ", new_input)
|
570 |
streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
|
571 |
generate_kwargs = dict(
|
@@ -583,18 +525,28 @@ def chat_response_stream(message: str, history: list, state: dict):
|
|
583 |
)
|
584 |
t = Thread(target=model.generate, kwargs=generate_kwargs)
|
585 |
t.start()
|
586 |
-
|
587 |
full_output = ""
|
588 |
for text in streamer:
|
589 |
full_output += text
|
590 |
time.sleep(0.05)
|
591 |
yield full_output
|
592 |
-
|
593 |
state["compressed_cache"] = past_key_values
|
594 |
return full_output
|
595 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
596 |
##########################################################################
|
597 |
-
# Gradio Interface
|
598 |
##########################################################################
|
599 |
CSS = """
|
600 |
body {
|
@@ -694,39 +646,57 @@ with gr.Blocks(css=CSS, theme=gr.themes.Soft()) as demo:
|
|
694 |
token_count_text = gr.Markdown("Number of tokens before compression: ")
|
695 |
retrieval_slider = gr.Slider(label="Select Compression Rate", minimum=1, maximum=32, step=1, value=2)
|
696 |
retrieval_info_text = gr.Markdown("Number of tokens after compression: ")
|
|
|
|
|
697 |
global_local_slider = gr.Radio(label="Global vs Local (0 is all RAG, 100 is all global)",
|
698 |
choices=["0%", "25%", "50%", "75%", "100%"], value="75%")
|
699 |
compress_button = gr.Button("Compress Document", interactive=False, elem_classes="upload-button")
|
|
|
|
|
700 |
|
701 |
file_input.change(
|
702 |
fn=auto_convert,
|
703 |
inputs=[file_input, url_input, do_ocr, do_table],
|
704 |
-
outputs=[markdown_output, token_count_text, retrieval_slider, retrieval_info_text, hidden_token_count, compress_button, compression_done, compressed_doc_state]
|
705 |
)
|
706 |
url_input.change(
|
707 |
fn=auto_convert,
|
708 |
inputs=[file_input, url_input, do_ocr, do_table],
|
709 |
-
outputs=[markdown_output, token_count_text, retrieval_slider, retrieval_info_text, hidden_token_count, compress_button, compression_done, compressed_doc_state]
|
710 |
)
|
711 |
do_ocr.change(
|
712 |
fn=auto_convert,
|
713 |
inputs=[file_input, url_input, do_ocr, do_table],
|
714 |
-
outputs=[markdown_output, token_count_text, retrieval_slider, retrieval_info_text, hidden_token_count, compress_button, compression_done, compressed_doc_state]
|
715 |
)
|
716 |
do_table.change(
|
717 |
fn=auto_convert,
|
718 |
inputs=[file_input, url_input, do_ocr, do_table],
|
719 |
-
outputs=[markdown_output, token_count_text, retrieval_slider, retrieval_info_text, hidden_token_count, compress_button, compression_done, compressed_doc_state]
|
720 |
)
|
721 |
retrieval_slider.change(
|
722 |
fn=update_retrieval_context,
|
723 |
inputs=[hidden_token_count, retrieval_slider],
|
724 |
outputs=retrieval_info_text
|
725 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
726 |
compress_button.click(
|
727 |
fn=prepare_compression_and_rag,
|
728 |
inputs=[markdown_output, retrieval_slider, global_local_slider, task_description_input, few_shot_input, compressed_doc_state],
|
729 |
outputs=[compressed_doc_state, compression_done]
|
|
|
|
|
|
|
730 |
)
|
731 |
|
732 |
with gr.Column(elem_classes="chatbot-container"):
|
|
|
3 |
import os
|
4 |
import time
|
5 |
from threading import Thread
|
6 |
+
import uuid
|
7 |
|
8 |
import gradio as gr
|
9 |
import spaces
|
|
|
19 |
from langchain_text_splitters import RecursiveCharacterTextSplitter
|
20 |
from transformers import AutoModelForCausalLM, AutoTokenizer, DynamicCache, TextIteratorStreamer
|
21 |
from transformers.models.llama.modeling_llama import rotate_half
|
|
|
22 |
|
23 |
from utils import (
|
24 |
calculate_tokens_suggest_compression_ratio,
|
|
|
26 |
update_retrieval_context,
|
27 |
)
|
28 |
|
|
|
|
|
29 |
# Initialize the model and tokenizer.
|
30 |
api_token = os.getenv("HUGGING_FACE_HUB_TOKEN")
|
31 |
model_name = "meta-llama/Llama-3.1-8B-Instruct"
|
|
|
35 |
model = model.eval()
|
36 |
model.to(device)
|
37 |
embedding_model = HuggingFaceBgeEmbeddings(
|
38 |
+
model_name="BAAI/bge-large-en-v1.5",
|
39 |
+
model_kwargs={"device": str(device)},
|
40 |
+
encode_kwargs={"normalize_embeddings": True},
|
41 |
+
query_instruction=""
|
42 |
+
)
|
|
|
43 |
|
44 |
# Create a chat template and split into prefix and suffix.
|
45 |
content_system = ""
|
|
|
118 |
self._seen_tokens = new_length
|
119 |
|
120 |
def save(self, path: str):
|
|
|
121 |
try:
|
122 |
os.makedirs(os.path.dirname(path), exist_ok=True)
|
123 |
torch.save(
|
|
|
129 |
|
130 |
@classmethod
|
131 |
def load(cls, path: str, device: str = "cpu") -> "FinchCache":
|
|
|
132 |
data = torch.load(path, map_location=device)
|
133 |
cache = cls()
|
134 |
cache.key_cache = [k.to(device) for k in data["key_cache"]]
|
|
|
136 |
cache._seen_tokens = cache.value_cache[0].size(2) if cache.value_cache else 0
|
137 |
return cache
|
138 |
|
|
|
|
|
139 |
def convert_to_markdown(file_objs, url, do_ocr, do_table_structure):
|
140 |
file_path = file_objs if file_objs is not None else url
|
141 |
pipeline_options = PdfPipelineOptions()
|
|
|
147 |
)
|
148 |
doc_converter = DocumentConverter(
|
149 |
allowed_formats=[InputFormat.PDF],
|
150 |
+
format_options={InputFormat.PDF: pdf_format_options}
|
|
|
|
|
151 |
)
|
|
|
|
|
152 |
loader = DoclingLoader(
|
153 |
file_path=file_path,
|
154 |
export_type=ExportType.MARKDOWN,
|
|
|
157 |
docs = loader.load()
|
158 |
return docs[0].page_content
|
159 |
|
|
|
160 |
def create_rag_index(collection_name, text_no_prefix):
|
|
|
161 |
text_splitter = RecursiveCharacterTextSplitter.from_huggingface_tokenizer(
|
162 |
+
tokenizer,
|
163 |
+
chunk_size=256,
|
164 |
+
chunk_overlap=0,
|
165 |
+
add_start_index=True,
|
166 |
+
strip_whitespace=True,
|
167 |
+
separators=["\n\n", "\n", ".", " ", ""],
|
168 |
+
)
|
|
|
169 |
docs = [Document(page_content=x) for x in text_splitter.split_text(text_no_prefix)]
|
170 |
vectorstore = Chroma.from_documents(collection_name=collection_name, persist_directory="./chroma_db", documents=docs, embedding=embedding_model)
|
171 |
return vectorstore
|
172 |
|
|
|
173 |
@spaces.GPU
|
174 |
def auto_convert(file_objs, url, do_ocr, do_table_structure):
|
175 |
+
# When a new file/URL is loaded, disable chat (compression not done)
|
176 |
+
chat_status = "Document not compressed yet. Please compress the document to enable chat."
|
177 |
if file_objs is None and (url is None or url.strip() == ""):
|
178 |
return (
|
179 |
gr.update(value=""),
|
180 |
"Number of tokens before compression: ",
|
181 |
+
gr.update(),
|
182 |
"Number of tokens after compression: ",
|
183 |
0,
|
184 |
+
gr.update(interactive=False),
|
185 |
False,
|
186 |
+
{},
|
187 |
+
chat_status
|
188 |
)
|
|
|
189 |
print("Converting to markdown")
|
190 |
+
try:
|
191 |
+
markdown = convert_to_markdown(file_objs, url, do_ocr, do_table_structure)
|
192 |
+
except Exception as e:
|
193 |
+
print("Error converting to markdown:", e)
|
194 |
+
return (
|
195 |
+
gr.update(value="Error converting document to markdown. Please try uploading another document format."),
|
196 |
+
"Number of tokens before compression: ",
|
197 |
+
gr.update(),
|
198 |
+
"Number of tokens after compression: ",
|
199 |
+
0,
|
200 |
+
gr.update(interactive=False),
|
201 |
+
False,
|
202 |
+
{},
|
203 |
+
chat_status
|
204 |
+
)
|
205 |
print("Done")
|
206 |
combined_text = prefix + markdown
|
207 |
print("Suggestioning Compression ratio")
|
|
|
214 |
token_count_str = f"Number of tokens before compression: {token_count}"
|
215 |
retrieval_str = f"Number of tokens after compression: {retrieval_tokens}"
|
216 |
slider_update = gr.update(value=default_ratio, minimum=min_ratio, maximum=max_ratio, step=1)
|
|
|
|
|
217 |
if combined_text.startswith(prefix):
|
218 |
rag_text = combined_text[len(prefix):]
|
219 |
else:
|
|
|
222 |
rag_index = create_rag_index(collection_name, rag_text)
|
223 |
state = {"rag_index": collection_name}
|
224 |
print("Done")
|
|
|
225 |
return (
|
226 |
+
combined_text,
|
227 |
+
token_count_str,
|
228 |
+
slider_update,
|
229 |
+
retrieval_str,
|
230 |
+
token_count,
|
231 |
+
gr.update(interactive=True), # Enable compress button if conversion succeeds.
|
232 |
False,
|
233 |
+
state,
|
234 |
+
chat_status
|
235 |
)
|
|
|
236 |
|
237 |
def get_compressed_kv_cache(sink_tokens, step_size, target_token_size, context_ids, context_attention_mask, question_ids, question_attention_mask):
|
238 |
device = model.device
|
|
|
248 |
max_context_tokens_allowed = model.config.max_position_embeddings - question_len
|
249 |
if total_len > max_context_tokens_allowed:
|
250 |
num_chunks = max(step_size, math.ceil(total_len / max_context_tokens_allowed))
|
|
|
251 |
if total_len <= sink_tokens or num_chunks == 1:
|
|
|
252 |
context_ids_list = [context_ids]
|
253 |
context_attention_mask_list = [context_attention_mask]
|
254 |
else:
|
|
|
255 |
remainder_len = total_len - sink_tokens
|
|
|
|
|
256 |
base = remainder_len // num_chunks
|
257 |
leftover = remainder_len % num_chunks
|
|
|
|
|
|
|
258 |
chunk_sizes = [sink_tokens + base]
|
|
|
|
|
259 |
for _ in range(num_chunks - 2):
|
260 |
chunk_sizes.append(base)
|
|
|
|
|
261 |
if num_chunks > 1:
|
262 |
chunk_sizes.append(base + leftover)
|
|
|
|
|
263 |
context_ids_list = []
|
264 |
context_attention_mask_list = []
|
265 |
offset = 0
|
|
|
268 |
context_ids_list.append(context_ids[:, offset:end])
|
269 |
context_attention_mask_list.append(context_attention_mask[:, offset:end])
|
270 |
offset = end
|
|
|
|
|
271 |
len_rest = max(total_len - sink_tokens, 1)
|
272 |
compression_factor = len_rest // target_token_size
|
273 |
if compression_factor < 1:
|
274 |
compression_factor = 1
|
|
|
275 |
tokenized_doc_chunks = []
|
276 |
for ids_chunk, mask_chunk in zip(context_ids_list, context_attention_mask_list):
|
277 |
tokenized_doc_chunks.append({"input_ids": ids_chunk, "attention_mask": mask_chunk})
|
|
|
278 |
print("Number of chunks: ", len(tokenized_doc_chunks))
|
|
|
279 |
rotary_emb = model.model.rotary_emb.to(device)
|
280 |
inv_freq = rotary_emb.inv_freq
|
281 |
batch_size = question_ids.size(0)
|
282 |
ones_mask = torch.ones(batch_size, 1, dtype=question_attention_mask.dtype, device=device)
|
|
|
283 |
cache = FinchCache()
|
284 |
past_cache_len = 0
|
285 |
past_attention_mask = torch.zeros(batch_size, 0, dtype=question_attention_mask.dtype, device=device)
|
286 |
num_chunks = len(tokenized_doc_chunks)
|
|
|
|
|
287 |
query_context_matrices = {}
|
|
|
|
|
288 |
def query_hook_fn(module, input, output):
|
289 |
layer_idx = getattr(module, "layer_idx", None)
|
290 |
if layer_idx is not None:
|
|
|
297 |
.transpose(1, 2)
|
298 |
.contiguous()
|
299 |
)
|
|
|
300 |
query_context_matrices[layer_idx] = query_states[:, :, _current_chunk_offset:, :].clone()
|
|
|
|
|
301 |
hooks = []
|
302 |
for i, layer in enumerate(model.model.layers):
|
303 |
+
layer.self_attn.q_proj.layer_idx = i
|
304 |
layer.self_attn.q_proj.num_query_heads = layer.self_attn.config.num_attention_heads
|
305 |
hook = layer.self_attn.q_proj.register_forward_hook(query_hook_fn)
|
306 |
hooks.append(hook)
|
|
|
|
|
307 |
for j, tokenized_doc_chunk in enumerate(tokenized_doc_chunks):
|
308 |
current_seq_length = tokenized_doc_chunk["input_ids"].size(1)
|
|
|
309 |
_current_chunk_offset = current_seq_length
|
|
|
310 |
query_context_matrices.clear()
|
|
|
|
|
311 |
chunk_input_ids = tokenized_doc_chunk["input_ids"].contiguous()
|
312 |
chunk_attention_mask = tokenized_doc_chunk["attention_mask"].contiguous()
|
313 |
segment_attention_mask = torch.cat(
|
|
|
315 |
).contiguous()
|
316 |
current_input_ids = torch.cat([chunk_input_ids, question_ids], dim=-1).contiguous()
|
317 |
current_attention_mask = torch.cat([segment_attention_mask, question_attention_mask], dim=-1).contiguous()
|
|
|
318 |
past_seen_tokens = cache.get_seq_length() if cache is not None else 0
|
319 |
cache_position = torch.arange(
|
320 |
past_seen_tokens + chunk_input_ids.shape[1],
|
|
|
330 |
cache_position=cache_position,
|
331 |
batch_size=current_input_ids.size(0),
|
332 |
).contiguous()
|
|
|
333 |
with torch.no_grad():
|
334 |
outputs = model.model(
|
335 |
input_ids=current_input_ids,
|
|
|
337 |
past_key_values=cache,
|
338 |
)
|
339 |
cache = outputs.past_key_values
|
|
|
340 |
len_question = question_ids.size(1)
|
|
|
341 |
for layer_idx in range(len(model.model.layers)):
|
342 |
key_matrix = cache.key_cache[layer_idx]
|
343 |
query_matrix = query_context_matrices[layer_idx]
|
|
|
353 |
query_matrix = (query_matrix * cos) + (rotate_half(query_matrix) * sin)
|
354 |
num_repeats = model.config.num_attention_heads // model.config.num_key_value_heads
|
355 |
key_matrix = repeat_kv(key_matrix, num_repeats)
|
|
|
356 |
scaling = math.sqrt(model.config.head_dim)
|
357 |
attention_matrix = torch.matmul(query_matrix, key_matrix.transpose(2, 3)) / scaling
|
358 |
causal_mask_sliced = causal_mask[:, :, :, : key_matrix.shape[-2]]
|
359 |
attention_matrix = attention_matrix + causal_mask_sliced
|
360 |
attention_matrix = torch.nn.functional.softmax(attention_matrix, dim=-1, dtype=torch.float32).to(query_matrix.dtype)
|
|
|
361 |
tol = 1e-8
|
362 |
binary_mask = (torch.abs(causal_mask_sliced.to(torch.float32)) < tol).to(torch.float32)
|
363 |
non_zero_counts = binary_mask.sum(dim=3, keepdim=True)
|
|
|
382 |
to_keep_new = int(current_seq_length // compression_factor)
|
383 |
k = min(past_cache_len + to_keep_new, full_context_size)
|
384 |
else:
|
385 |
+
desired_final = sink_tokens + target_token_size + len_question
|
386 |
k = desired_final if full_context_size >= desired_final else full_context_size
|
387 |
k = max(k, sink_tokens)
|
388 |
selected_indices = torch.topk(attention_matrix, k, dim=-1).indices
|
389 |
selected_indices, _ = torch.sort(selected_indices, dim=-1)
|
390 |
cache.compress_cache(layer_idx, selected_indices, inv_freq)
|
|
|
391 |
past_cache_len = cache._seen_tokens
|
392 |
past_attention_mask = torch.ones(1, past_cache_len, device=device)
|
|
|
|
|
393 |
for hook in hooks:
|
394 |
hook.remove()
|
|
|
395 |
return cache
|
396 |
|
|
|
397 |
def run_naive_rag_query(collection_name, query, rag_token_size, prefix, task, few_shot_examples):
|
|
|
|
|
|
|
|
|
398 |
k = max(1, rag_token_size // 256)
|
399 |
+
vectorstore = Chroma(persist_directory="./chroma_db", embedding_function=embedding_model, collection_name=collection_name)
|
400 |
retriever = vectorstore.as_retriever(search_type="similarity", search_kwargs={"k": k})
|
401 |
retrieved_docs = retriever.invoke(query)
|
402 |
for doc in retrieved_docs:
|
|
|
404 |
print(doc.page_content)
|
405 |
print("=================")
|
406 |
formatted_context = "\n\n".join([doc.page_content for doc in retrieved_docs])
|
|
|
407 |
rag_context = prefix + "Retrieved context: \n" + formatted_context + task + few_shot_examples
|
|
|
408 |
return rag_context
|
409 |
|
|
|
410 |
@spaces.GPU
|
411 |
def prepare_compression_and_rag(combined_text, retrieval_slider_value, global_local_value, task_description, few_shot, state):
|
|
|
|
|
|
|
412 |
percentage = int(global_local_value.replace('%', ''))
|
413 |
question_text = task_description + "\n" + few_shot
|
414 |
context_encoding = tokenizer(combined_text, return_tensors="pt").to(device)
|
|
|
418 |
question_ids = question_encoding["input_ids"]
|
419 |
question_attention_mask = question_encoding["attention_mask"]
|
420 |
retrieval_context_length = int(context_ids.size(1) / retrieval_slider_value)
|
421 |
+
# Compute token breakdown for display (KV compress vs RAG tokens)
|
422 |
+
rag_tokens = int(retrieval_context_length * (1.0 - (percentage / 100)))
|
423 |
+
kv_tokens = retrieval_context_length - rag_tokens
|
424 |
+
print(f"KV Compress Tokens: {kv_tokens}, RAG Tokens: {rag_tokens}")
|
425 |
if percentage > 0:
|
426 |
target_token_size = int(retrieval_context_length * (percentage / 100))
|
427 |
print("Target token size for compression: ", target_token_size)
|
428 |
step_size = 2
|
429 |
start_time_prefill = time.perf_counter()
|
430 |
+
try:
|
431 |
+
past_key_values = copy.deepcopy(get_compressed_kv_cache(sink_tokens, step_size, target_token_size,
|
432 |
+
context_ids, context_attention_mask,
|
433 |
+
question_ids, question_attention_mask))
|
434 |
+
except Exception as e:
|
435 |
+
print("Error during KV cache compression:", e)
|
436 |
+
state["error"] = "Error during KV cache compression. Please try lowering the compression ratio and try again."
|
437 |
+
return state, False
|
438 |
compressed_length = past_key_values.get_seq_length()
|
439 |
print("Context size after compression: ", compressed_length)
|
440 |
print("Compression rate: ", context_ids.size(1) / compressed_length)
|
|
|
443 |
target_token_size = 0
|
444 |
past_key_values = FinchCache()
|
445 |
compressed_length = past_key_values.get_seq_length()
|
|
|
|
|
446 |
cache_name = "default_cache_" + uuid.uuid4().hex[:6] + ".pt"
|
447 |
save_dir = "./cache_dir"
|
448 |
os.makedirs(save_dir, exist_ok=True)
|
449 |
save_path = os.path.join(save_dir, cache_name)
|
450 |
past_key_values.save(save_path)
|
|
|
|
|
451 |
collection_name = state.get("rag_index", None)
|
452 |
if collection_name is None:
|
453 |
print("Collection name not found creating a new one.")
|
|
|
457 |
rag_text = combined_text
|
458 |
collection_name = "default_collection_" + uuid.uuid4().hex[:6]
|
459 |
rag_index = create_rag_index(collection_name, rag_text)
|
|
|
460 |
state.update({
|
461 |
"compressed_cache": save_path,
|
462 |
"compressed_length": compressed_length,
|
|
|
467 |
"task_description": task_description,
|
468 |
"few_shot": few_shot,
|
469 |
"retrieval_slider": retrieval_context_length,
|
470 |
+
"prefill_time": time.perf_counter() - start_time_prefill,
|
471 |
+
"compression_done": True,
|
472 |
+
"tokens_breakdown": f"KV Compress Tokens: {kv_tokens}, RAG Tokens: {rag_tokens}",
|
473 |
+
"chat_feedback": "Document compressed successfully. You can now chat."
|
474 |
})
|
475 |
return state, True
|
476 |
|
|
|
477 |
@spaces.GPU
|
478 |
def chat_response_stream(message: str, history: list, state: dict):
|
479 |
+
# Check if the document is compressed before allowing chat
|
480 |
+
if not state.get("compression_done", False) or "compressed_cache" not in state:
|
481 |
+
yield "Document not compressed yet. Please compress the document first to enable chat."
|
482 |
+
return
|
483 |
user_message = message
|
484 |
save_path = state["compressed_cache"]
|
485 |
past_key_values = FinchCache.load(save_path, device=model.device)
|
|
|
|
|
|
|
|
|
486 |
compressed_length = past_key_values.get_seq_length()
|
487 |
collection_name = state["rag_index"]
|
488 |
retrieval_slider_value = state["retrieval_slider"]
|
489 |
percentage = state["global_local"]
|
|
|
490 |
rag_retrieval_size = int(retrieval_slider_value * (1.0 - (percentage / 100)))
|
491 |
print("RAG retrieval size: ", rag_retrieval_size)
|
|
|
492 |
if percentage == 0:
|
493 |
rag_prefix = prefix
|
494 |
rag_task = state["task_description"]
|
|
|
508 |
eos_block = torch.full((1, compressed_length), tokenizer.eos_token_id, device=device, dtype=torch.long)
|
509 |
new_input_ids = torch.cat([eos_block, tokenized_new_input["input_ids"]], dim=-1)
|
510 |
new_attention_mask = torch.cat([torch.ones((1, compressed_length), device=device), tokenized_new_input["attention_mask"]], dim=-1)
|
|
|
511 |
print("New input is: ", new_input)
|
512 |
streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
|
513 |
generate_kwargs = dict(
|
|
|
525 |
)
|
526 |
t = Thread(target=model.generate, kwargs=generate_kwargs)
|
527 |
t.start()
|
|
|
528 |
full_output = ""
|
529 |
for text in streamer:
|
530 |
full_output += text
|
531 |
time.sleep(0.05)
|
532 |
yield full_output
|
|
|
533 |
state["compressed_cache"] = past_key_values
|
534 |
return full_output
|
535 |
|
536 |
+
def update_token_breakdown(token_count, retrieval_slider_value, global_local_value):
|
537 |
+
try:
|
538 |
+
token_count = int(token_count)
|
539 |
+
slider_val = float(retrieval_slider_value)
|
540 |
+
percentage = int(global_local_value.replace('%', ''))
|
541 |
+
retrieval_context_length = int(token_count / slider_val)
|
542 |
+
rag_tokens = int(retrieval_context_length * (1 - (percentage / 100)))
|
543 |
+
kv_tokens = retrieval_context_length - rag_tokens
|
544 |
+
return f"KV Compress Tokens: {kv_tokens}, RAG Tokens: {rag_tokens}"
|
545 |
+
except Exception as e:
|
546 |
+
return "Token breakdown unavailable."
|
547 |
+
|
548 |
##########################################################################
|
549 |
+
# Gradio Interface
|
550 |
##########################################################################
|
551 |
CSS = """
|
552 |
body {
|
|
|
646 |
token_count_text = gr.Markdown("Number of tokens before compression: ")
|
647 |
retrieval_slider = gr.Slider(label="Select Compression Rate", minimum=1, maximum=32, step=1, value=2)
|
648 |
retrieval_info_text = gr.Markdown("Number of tokens after compression: ")
|
649 |
+
# New widget for token breakdown (KV vs RAG)
|
650 |
+
tokens_breakdown_text = gr.Markdown("Token breakdown will appear here.")
|
651 |
global_local_slider = gr.Radio(label="Global vs Local (0 is all RAG, 100 is all global)",
|
652 |
choices=["0%", "25%", "50%", "75%", "100%"], value="75%")
|
653 |
compress_button = gr.Button("Compress Document", interactive=False, elem_classes="upload-button")
|
654 |
+
# New widget for chat status feedback
|
655 |
+
chat_status_text = gr.Markdown("Document not compressed yet. Please compress the document to enable chat.")
|
656 |
|
657 |
file_input.change(
|
658 |
fn=auto_convert,
|
659 |
inputs=[file_input, url_input, do_ocr, do_table],
|
660 |
+
outputs=[markdown_output, token_count_text, retrieval_slider, retrieval_info_text, hidden_token_count, compress_button, compression_done, compressed_doc_state, chat_status_text]
|
661 |
)
|
662 |
url_input.change(
|
663 |
fn=auto_convert,
|
664 |
inputs=[file_input, url_input, do_ocr, do_table],
|
665 |
+
outputs=[markdown_output, token_count_text, retrieval_slider, retrieval_info_text, hidden_token_count, compress_button, compression_done, compressed_doc_state, chat_status_text]
|
666 |
)
|
667 |
do_ocr.change(
|
668 |
fn=auto_convert,
|
669 |
inputs=[file_input, url_input, do_ocr, do_table],
|
670 |
+
outputs=[markdown_output, token_count_text, retrieval_slider, retrieval_info_text, hidden_token_count, compress_button, compression_done, compressed_doc_state, chat_status_text]
|
671 |
)
|
672 |
do_table.change(
|
673 |
fn=auto_convert,
|
674 |
inputs=[file_input, url_input, do_ocr, do_table],
|
675 |
+
outputs=[markdown_output, token_count_text, retrieval_slider, retrieval_info_text, hidden_token_count, compress_button, compression_done, compressed_doc_state, chat_status_text]
|
676 |
)
|
677 |
retrieval_slider.change(
|
678 |
fn=update_retrieval_context,
|
679 |
inputs=[hidden_token_count, retrieval_slider],
|
680 |
outputs=retrieval_info_text
|
681 |
)
|
682 |
+
# Update token breakdown when slider or global/local changes
|
683 |
+
retrieval_slider.change(
|
684 |
+
fn=update_token_breakdown,
|
685 |
+
inputs=[hidden_token_count, retrieval_slider, global_local_slider],
|
686 |
+
outputs=tokens_breakdown_text
|
687 |
+
)
|
688 |
+
global_local_slider.change(
|
689 |
+
fn=update_token_breakdown,
|
690 |
+
inputs=[hidden_token_count, retrieval_slider, global_local_slider],
|
691 |
+
outputs=tokens_breakdown_text
|
692 |
+
)
|
693 |
compress_button.click(
|
694 |
fn=prepare_compression_and_rag,
|
695 |
inputs=[markdown_output, retrieval_slider, global_local_slider, task_description_input, few_shot_input, compressed_doc_state],
|
696 |
outputs=[compressed_doc_state, compression_done]
|
697 |
+
).then(
|
698 |
+
fn=lambda state: gr.update(value="Document compressed successfully. You can now chat."),
|
699 |
+
outputs=chat_status_text
|
700 |
)
|
701 |
|
702 |
with gr.Column(elem_classes="chatbot-container"):
|