erik-svensson-cm's picture
Update handler.py
8b38935 verified
import gc
import logging
import re
import time
from collections import Counter
from typing import List, Optional
import torch
from pydantic import BaseModel, ValidationError
from transformers import (
AutoTokenizer,
AutoModelForSequenceClassification,
MarianTokenizer,
MarianMTModel,
pipeline,
PreTrainedTokenizer,
AutoModelForCausalLM, BitsAndBytesConfig
)
# Set up logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
B_INST, E_INST = "[INST] ", " [/INST]"
MAX_SUMMARY_TOKENS = 15900
MAX_TRANSLATION_TOKENS = 256
MAX_SENTIMENT_TOKENS = 256
class InferenceConfig(BaseModel):
sentiment: bool = False
summarize: bool = False
prompt: Optional[str] = None
def split_by_token_limit_with_speakers(text: str, max_tokens: int, tokenizer: PreTrainedTokenizer) -> List[str]:
"""
Split transcript text into chunks with max token length, preserving speaker turns.
"""
# Regex to extract speaker turns
speaker_pattern = re.compile(r"(Agent:|Customer:)")
parts = speaker_pattern.split(text)
if not parts[0].strip():
parts = parts[1:] # remove empty prefix if present
# Pair up speaker labels with their content
turns = [(parts[i], parts[i + 1].strip()) for i in range(0, len(parts), 2)]
chunks = []
current_chunk = ""
current_token_count = 0
for speaker, utterance in turns:
turn_text = f"{speaker} {utterance}"
tokenized = tokenizer(turn_text, add_special_tokens=False, return_attention_mask=True)
turn_tokens = len(tokenized["input_ids"])
# If turn fits, add to current chunk
if current_token_count + turn_tokens <= max_tokens:
if current_chunk:
current_chunk += " " + turn_text
else:
current_chunk = turn_text
current_token_count += turn_tokens
else:
# If turn itself is too long, split by words
if turn_tokens > max_tokens:
words = utterance.split()
partial = []
for word in words:
candidate = f"{speaker} {' '.join(partial + [word])}"
tokens = len(tokenizer(candidate, add_special_tokens=False)["input_ids"])
if tokens > max_tokens:
# finalize current partial
if partial:
chunks.append(f"{speaker} {' '.join(partial)}")
partial = [word]
else:
partial.append(word)
if partial:
chunks.append(f"{speaker} {' '.join(partial)}")
current_chunk = ""
current_token_count = 0
else:
# finalize current chunk and start a new one
if current_chunk:
chunks.append(current_chunk)
current_chunk = turn_text
current_token_count = turn_tokens
if current_chunk:
chunks.append(current_chunk)
return chunks
def translate_chunks(chunks: list[str], tokenizer: MarianTokenizer, model: MarianMTModel, device: str = "cpu") -> list[str]:
model.eval()
translated_chunks = []
start_time = time.time()
for i in range(0, len(chunks), 8):
batch = chunks[i:i + 8]
inputs = tokenizer(
batch,
return_tensors="pt",
padding=True,
truncation=True,
max_length=512
).to(device)
with torch.no_grad():
with torch.amp.autocast("cuda"):
translated = model.generate(**inputs)
translated_texts = tokenizer.batch_decode(
translated,
skip_special_tokens=True
)
translated_chunks.extend(translated_texts)
gen_time = time.time() - start_time
logger.info(f"Generated translation in {gen_time:.2f}s.")
return translated_chunks
def summarize_with_mistral(text: str, model, tokenizer, prompt_helper: Optional[str] = None) -> str:
"""
Summarize text using the Mistral summarization model.
"""
# Format prompt according to model's requirements
B_INST, E_INST = "[INST] ", " [/INST]"
if prompt_helper is None:
prompt_helper = "You are given a transcript of a phone call between a customer and a company agent. The agent is always the one who initiates the call. Summarize the conversation clearly and concisely for an internal admin reader. Focus on the purpose of the call, key discussion points, any actions taken or agreed upon, and any relevant customer sentiment or concerns."
prompt = f"""{B_INST}{prompt_helper}\n\n[TEXT_START]\n\n{text}\n\n[TEXT_END]\n\n{E_INST}"""
# Tokenize with attention mask
inputs = tokenizer(
prompt,
return_tensors="pt",
truncation=True,
max_length=16000,
return_attention_mask=True
).to(model.device)
logger.info(f'Length of input tokens: {len(inputs["input_ids"][0])}')
# Move inputs to the same device as model
inputs = {k: v.to(model.device) for k, v in inputs.items()}
with torch.no_grad():
with torch.amp.autocast("cuda"):
outputs = model.generate(
**inputs,
max_new_tokens=256,
temperature=0.3,
top_p=0.9,
top_k=20,
repetition_penalty=1.2,
do_sample=True,
pad_token_id=tokenizer.eos_token_id
)
# Extract only the generated summary (remove the prompt)
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
summary = generated_text.split("[/INST]")[-1].strip()
return summary
def handle_long_summarization(text: str, tokenizer, model) -> str:
"""Handle texts longer than context window using recursive summarization"""
chunks, _ = split_by_token_limit_with_speakers(
text,
MAX_SUMMARY_TOKENS / 2, # Use half context for chunking
tokenizer
)
logger.info(f"Processing {len(chunks)} chunks for recursive summarization")
summaries = []
for i, chunk in enumerate(chunks):
logger.info(f"Summarizing chunk {i + 1}/{len(chunks)}")
summary = summarize_with_mistral(
chunk,
model,
tokenizer
)
summaries.append(summary)
combined = " ".join(summaries)
# Final summary if still too long
if len(tokenizer(combined)["input_ids"]) > MAX_SUMMARY_TOKENS:
return handle_long_summarization(combined, model, tokenizer)
return summarize_with_mistral(
combined,
model,
tokenizer
)
class EndpointHandler():
def __init__(self, path=""):
# Determine device
self.device = "cuda" if torch.cuda.is_available() else "cpu"
logger.info(f"Using device: {self.device}")
# Initialize sentiment analysis pipeline
self.sentiment_pipeline = None
self.sentiment_tokenizer = None
try:
self.sentiment_tokenizer = AutoTokenizer.from_pretrained("KBLab/megatron-bert-large-swedish-cased-165k")
sentiment_model = AutoModelForSequenceClassification.from_pretrained(
"KBLab/robust-swedish-sentiment-multiclass")
self.sentiment_pipeline = pipeline(
"sentiment-analysis",
model=sentiment_model,
tokenizer=self.sentiment_tokenizer,
device=self.device
)
logger.info("Sentiment analysis pipeline initialized")
except Exception as e:
logger.error(f"Error initializing sentiment pipeline: {e}")
# Initialize Swedish to English translation
self.model_sv_en = None
self.tokenizer_sv_en = None
try:
model_name_sv_en = "Helsinki-NLP/opus-mt-sv-en"
self.tokenizer_sv_en = MarianTokenizer.from_pretrained(model_name_sv_en)
self.model_sv_en = MarianMTModel.from_pretrained(model_name_sv_en)
self.model_sv_en.to(self.device)
logger.info("Swedish to English translation model initialized")
except Exception as e:
logger.error(f"Error initializing sv-en translation: {e}")
# Initialize English to Swedish translation
self.model_en_sv = None
self.tokenizer_en_sv = None
try:
model_name_en_sv = "Helsinki-NLP/opus-mt-en-sv"
self.tokenizer_en_sv = MarianTokenizer.from_pretrained(model_name_en_sv)
self.model_en_sv = MarianMTModel.from_pretrained(model_name_en_sv)
self.model_en_sv.to(self.device)
logger.info("English to Swedish translation model initialized")
except Exception as e:
logger.error(f"Error initializing en-sv translation: {e}")
# Initialize summarization model (using simpler device management)
self.summarizer_model = None
self.summarizer_tokenizer = None
try:
model_name = "Trelis/Mistral-7B-Instruct-v0.1-Summarize-16k"
self.summarizer_tokenizer = AutoTokenizer.from_pretrained(model_name)
# Set pad token if not already set
if self.summarizer_tokenizer.pad_token is None:
self.summarizer_tokenizer.pad_token = self.summarizer_tokenizer.eos_token
# Load model with appropriate dtype
# Define the quantization config (4-bit in this case)
bnb_config = BitsAndBytesConfig(
load_in_4bit=True, # or load_in_8bit=True
bnb_4bit_compute_dtype="float16",
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4", # "nf4" is better than "fp4" in most cases
)
self.summarizer_model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.float16,
device_map="auto",
quantization_config=bnb_config,
)
logger.info("Summarization model initialized")
except Exception as e:
logger.error(f"Error initializing summarizer: {e}")
def __call__(self, inputs):
# Extract inputs and parameters
conversation = inputs.get("inputs", "")
parameters = inputs.get("parameters", {})
# Validate parameters
try:
config = InferenceConfig(**parameters)
except ValidationError as e:
logger.error(f"Error validating parameters: {e}")
return {"error": f"Error validating parameters: {e}"}
logger.info(f"Processing conversation with parameters: {config}")
# Initialize result dictionary
result = {}
# Perform sentiment analysis
if config.sentiment and self.sentiment_pipeline and self.sentiment_tokenizer:
logger.info("Performing sentiment analysis...")
try:
text_chunks_for_sentiment = split_by_token_limit_with_speakers(conversation, MAX_SENTIMENT_TOKENS,
self.sentiment_tokenizer)
sentiment_results = []
for i, chunk in enumerate(text_chunks_for_sentiment):
if chunk.strip(): # Skip empty chunks
try:
chunk_results = self.sentiment_pipeline(chunk)
if isinstance(chunk_results, list):
sentiment_results.extend(chunk_results)
else:
sentiment_results.append(chunk_results)
except Exception as chunk_error:
logger.warning(f"Failed to analyze sentiment for chunk {i}: {chunk_error}")
continue
if sentiment_results:
labels = [r['label'] for r in sentiment_results]
scores = [r['score'] for r in sentiment_results]
# Get most common sentiment
most_common_label = Counter(labels).most_common(1)[0][0]
avg_score = sum(scores) / len(scores)
result["sentiment"] = {
"label": most_common_label,
"score": round(avg_score, 2),
"details": {
"all_sentiments": labels,
"distribution": dict(Counter(labels))
}
}
else:
result["sentiment"] = {"label": "neutral", "score": 0.0}
torch.cuda.empty_cache()
del text_chunks_for_sentiment
gc.collect()
logger.info(f"Sentiment analysis completed: {result['sentiment']['label']}")
except Exception as e:
logger.error(f"Sentiment analysis error: {str(e)}")
result["sentiment"] = {"error": f"Sentiment analysis failed: {str(e)}"}
# Perform translation and summarization
if config.summarize and all([self.summarizer_model, self.model_sv_en, self.model_en_sv]):
logger.info("Performing translation and summarization...")
try:
# Translate Swedish to English for summarization
logger.info("Translating to English...")
text_chunks_for_translation = split_by_token_limit_with_speakers(
conversation, MAX_TRANSLATION_TOKENS, self.tokenizer_sv_en)
translated_chunks = translate_chunks(
text_chunks_for_translation,
self.tokenizer_sv_en,
self.model_sv_en,
device=self.device
)
# Combine translated chunks
translated_conversation = " ".join(translated_chunks)
logger.info(f"Translated conversation length: {len(translated_conversation)} chars")
# Generate summary using Mistral
logger.info("Generating summary with Mistral...")
# Handle long texts
token_count = len(self.summarizer_tokenizer(translated_conversation)["input_ids"])
torch.cuda.empty_cache()
gc.collect()
if token_count > MAX_SUMMARY_TOKENS:
logger.warning(f"Text too long ({token_count} tokens), using recursive summarization")
english_summary = handle_long_summarization(translated_conversation, self.summarizer_tokenizer, self.summarizer_model)
else:
english_summary = summarize_with_mistral(
translated_conversation,
self.summarizer_model,
self.summarizer_tokenizer,
)
torch.cuda.empty_cache()
del translated_conversation
gc.collect()
if english_summary:
logger.info(f"Generated English summary: {english_summary[:100]}...")
# Translate summary back to Swedish
logger.info("Translating summary back to Swedish...")
swedish_summary_list = translate_chunks(
[english_summary],
self.tokenizer_en_sv,
self.model_en_sv,
device=self.device
)
swedish_summary = swedish_summary_list[0] if swedish_summary_list else english_summary
result["summary"] = {
"swedish": swedish_summary,
"english": english_summary
}
else:
result["summary"] = {"error": "Failed to generate summary"}
logger.info("Summarization completed")
except Exception as e:
logger.error(f"Translation or summarization error: {str(e)}")
result["summary"] = {"error": f"Summarization failed: {str(e)}"}
torch.cuda.empty_cache()
gc.collect()
return result