Feature(LLMLingua): add KV-Cache Compression & HF Space Demo

Author: iofu728

README.md

@@ -2,20 +2,22 @@
 <img src="images/LLMLingua_logo.png" alt="LLMLingua" style="width: 20%; min-width: 100px; display: block; margin: auto;">
 </p>
 
-# LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models [[paper]()] & LongLLMLingua [[paper]()]
+# LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models [paper] & LongLLMLingua [paper]
 
 https://github.com/microsoft/LLMLingua/assets/30883354/ef52995c-ef3c-4eac-a9fd-1acb491c325b
 
+You can try the LLMLingua demo in [HF Space](https://huggingface.co/spaces/microsoft/LLMLingua).
+
 ## Tl;DR
 
 LLMLingua, that uses a well-trained small language model after alignment, such as GPT2-small or LLaMA-7B, to detect the unimportant tokens in the prompt and enable inference with the compressed prompt in black-box LLMs, achieving up to 20x compression with minimal performance loss.
 
-[LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models]() (EMNLP 2023).
+LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models (EMNLP 2023).<br>
 _Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang and Lili Qiu_
 
 LongLLMLingua is a method that enhances LLMs' ability to perceive key information in long-context scenarios using prompt compression, achieveing up to $28.5 in cost savings per 1,000 samples while also improving performance.
 
-[LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression]() (Under Review).
+LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression (Under Review).<br>
 _Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang and Lili Qiu_
 
 ## 🎥 Overview

llmlingua/prompt_compressor.py

@@ -24,12 +24,13 @@ def __init__(
         self.load_model(model_name, device_map, use_auth_token)
         self.sbert = None
         self.open_api_config = open_api_config
+        self.cache_bos_num = 10
 
     def load_model(
         self, model_name: str, device_map: str = "cuda", use_auth_token: bool = False
     ):
-        config = AutoConfig.from_pretrained(model_name)
-        tokenizer = AutoTokenizer.from_pretrained(model_name)
+        config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
+        tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
         tokenizer.padding_side = "left"
         tokenizer.pad_token_id = (
             config.pad_token_id if config.pad_token_id else tokenizer.eos_token_id
@@ -40,12 +41,11 @@ def load_model(
         if "cuda" in device_map or "cpu" in device_map:
             model = AutoModelForCausalLM.from_pretrained(
                 model_name,
-                torch_dtype="auto",
+                torch_dtype="auto" if device_map == "cuda" else torch.float32,
                 config=config,
                 ignore_mismatched_sizes=True,
+                trust_remote_code=True,
             ).to(device_map)
-            if device_map == "cpu":
-                model = model.type(torch.float32)
         else:
             model = AutoModelForCausalLM.from_pretrained(
                 model_name,
@@ -56,10 +56,12 @@ def load_model(
                 offload_state_dict=True,
                 cache_dir="/tmp/cache",
                 use_auth_token=use_auth_token,
+                trust_remote_code=True,
             )
         self.tokenizer = tokenizer
         self.model = model
         self.context_idxs = []
+        self.max_position_embeddings = config.max_position_embeddings
 
     def get_ppl(
         self,
@@ -83,7 +85,7 @@ def get_ppl(
             past_length = 0
         if end is None:
             end = input_ids.shape[1]
-        end = min(end, past_length + 4096)
+        end = min(end, past_length + self.max_position_embeddings)
         with torch.no_grad():
             response = self.model(
                 input_ids[:, past_length:end],
@@ -145,11 +147,17 @@ def compress_prompt(
         assert not (
             rank_method == "longllmlingua" and not question
         ), "In the LongLLMLingua, it is necessary to set a question."
+        if condition_compare and "_condition" not in condition_in_question:
+            condition_in_question += "_condition"
         if rank_method == "longllmlingua":
             if condition_in_question == "none":
                 condition_in_question = "after"
         elif rank_method == "llmlingua":
-            condition_in_question = "none"
+            condition_in_question = (
+                "none"
+                if "_condition" not in condition_in_question
+                else "none_condition"
+            )
         origin_tokens = len(
             encoding.encode("\n\n".join([instruction] + context + [question]).strip())
         )
@@ -653,8 +661,52 @@ def iterative_compress_prompt(
             keep_flag = torch.tensor(keep_flag).to(self.device)
         past_key_values, past_loss, ready_end = None, None, 0
         self_past_key_values, self_past_loss, self_ready_end = None, None, 0
+        pop_compressed_input_ids, pop_self_compressed_input_ids = None, None
         idx = 0
         while end <= compressed_input_ids.shape[1]:
+            if end > self.max_position_embeddings and past_key_values is not None:
+                # KV-Cache Compression
+                e, s = end - self.max_position_embeddings, self.cache_bos_num
+                if pop_compressed_input_ids is None:
+                    pop_compressed_input_ids = compressed_input_ids[:, :e]
+                else:
+                    pop_compressed_input_ids = torch.cat(
+                        [pop_compressed_input_ids, compressed_input_ids[:, :e]], dim=-1
+                    )
+                compressed_input_ids = compressed_input_ids[:, e:]
+                compressed_attention_mask = compressed_attention_mask[:, e:]
+                past_key_values = [
+                    [
+                        torch.cat([k[..., :s, :], k[..., s + e :, :]], dim=-2),
+                        torch.cat([v[..., :s, :], v[..., s + e :, :]], dim=-2),
+                    ]
+                    for k, v in past_key_values
+                ]
+                end, ready_end = end - e, ready_end - e
+                if condition_compare:
+                    self_ready_end -= e
+                    if pop_self_compressed_input_ids is None:
+                        pop_self_compressed_input_ids = self_compressed_input_ids[:, :e]
+                    else:
+                        pop_self_compressed_input_ids = torch.cat(
+                            [
+                                pop_self_compressed_input_ids,
+                                self_compressed_input_ids[:, :e],
+                            ],
+                            dim=-1,
+                        )
+                    self_compressed_input_ids = self_compressed_input_ids[:, e:]
+                    self_compressed_attention_mask = self_compressed_attention_mask[
+                        :, e:
+                    ]
+                    self_past_key_values = [
+                        [
+                            torch.cat([k[..., :s, :], k[..., s + e :, :]], dim=-2),
+                            torch.cat([v[..., :s, :], v[..., s + e :, :]], dim=-2),
+                        ]
+                        for k, v in self_past_key_values
+                    ]
+
             loss, past_key_values = self.get_ppl(
                 "",
                 "token",
@@ -762,6 +814,10 @@ def iterative_compress_prompt(
                 )
                 end += iterative_size
             idx += 1
+        if pop_compressed_input_ids is not None:
+            compressed_input_ids = torch.cat(
+                [pop_compressed_input_ids, compressed_input_ids], dim=-1
+            )
         return compressed_input_ids[:, start:], compressed_attention_mask[:, start:]
 
     def recover(

llmlingua/version.py

@@ -2,7 +2,7 @@
 _MINOR = "1"
 # On master and in a nightly release the patch should be one ahead of the last
 # released build.
-_PATCH = "1"
+_PATCH = "2"
 # This is mainly for nightly builds which have the suffix ".dev$DATE". See
 # https://semver.org/#is-v123-a-semantic-version for the semantics.
 _SUFFIX = ""