Add StreamingLLM for llamacpp & llamacpp_HF (2nd attempt) (oobabooga#…

…5669)

modules/cache_utils.py

@@ -0,0 +1,108 @@
+import torch
+
+from modules import shared
+from modules.logging_colors import logger
+
+
+def process_llamacpp_cache(model, new_sequence, past_sequence):
+    i1, i2, j1, j2 = find_longest_common_substring_indices(past_sequence, new_sequence)
+    overlap_length = i2 - i1 + 1
+
+    # Do StreamingLLM if i1 > 0 (ie the longest common subsequence is not a prefix)
+    # and the overlap length is sufficiently long.
+    if i1 > 0 and overlap_length > 0.2 * len(new_sequence):
+
+        new_sequence = torch.tensor(new_sequence)
+        past_sequence = torch.tensor(past_sequence)
+
+        prefix_length = find_prefix_length(past_sequence[:i1], new_sequence[:j1])
+        sink_length = prefix_length
+        if sink_length < shared.args.attention_sink_size:
+            sink_length = shared.args.attention_sink_size
+
+        removed_length = i1 - sink_length
+
+        matching_prefix = past_sequence[:prefix_length]
+        removed_chunk = past_sequence[sink_length:i1]
+        overlapping_sequence = new_sequence[j1:j2 + 1]
+        added_chunk = new_sequence[j2 + 1:]
+
+        # print(past_sequence)
+        # print(new_sequence)
+
+        print()
+        print('MATCHING PREFIX=', repr(shared.tokenizer.decode(matching_prefix)))
+        print('ADDED CHUNK=', repr(shared.tokenizer.decode(added_chunk)))
+        print('REMOVED CHUNK=', repr(shared.tokenizer.decode(removed_chunk)))
+        print()
+
+        # Remove interval [sink_length, sink_length + removed_length) from the context
+        # Subtract removed_length from model.n_tokens
+        model._ctx.kv_cache_seq_rm(0, sink_length, sink_length + removed_length)
+        model._ctx.kv_cache_seq_shift(0, sink_length + removed_length, -1, -removed_length)
+
+        new_sequence = new_sequence.tolist()
+        model.input_ids[:j2 + 1] = new_sequence[:j2 + 1]
+        model.n_tokens = j2 + 1
+
+        return new_sequence[:j2 + 1]
+    else:
+        return past_sequence
+
+
+def find_prefix_length(past_seq, seq_tensor):
+    '''
+    Given two torch tensors, finds the length of the longest
+    common prefix between the two.
+    '''
+    min_length = min(past_seq.shape[0], seq_tensor.shape[0])
+    indices = torch.nonzero(~torch.eq(past_seq[:min_length], seq_tensor[:min_length]))
+    if len(indices) > 0:
+        prefix_length = indices[0].item()
+    else:
+        prefix_length = min_length
+
+    return prefix_length
+
+
+def find_longest_common_substring_indices(list1, list2):
+    '''
+    Given two lists, solves the Longest Common Substring problem.
+
+    It returns the indices where the substring starts and ends in
+    s1 and s2.
+
+    Example:
+
+    ir, jr, ir2, jr2 = find_longest_common_substring_indices(s1, s2)
+    print(s1[ir:jr + 1])
+    print(s2[ir2:jr2 + 1])
+
+    Adapted from
+    https://rosettacode.org/wiki/Longest_common_substring#Python
+    '''
+
+    len_list1, len_list2 = len(list1), len(list2)
+    start_index_list1, end_index_list1 = 0, -1
+    start_index_list2, end_index_list2 = 0, -1
+
+    for index1 in range(len_list1):
+        try:
+            index2 = list2.index(list1[index1])
+        except ValueError:
+            continue
+        while index2 >= 0:
+            temp_index1, temp_index2 = index1, index2
+            while temp_index1 < len_list1 and temp_index2 < len_list2 and list2[temp_index2] == list1[temp_index1]:
+                if temp_index1 - index1 >= end_index_list1 - start_index_list1:
+                    start_index_list1, end_index_list1 = index1, temp_index1
+                    start_index_list2, end_index_list2 = index2, temp_index2
+
+                temp_index1 += 1
+                temp_index2 += 1
+            try:
+                index2 = list2.index(list1[index1], index2 + 1)
+            except ValueError:
+                break
+
+    return start_index_list1, end_index_list1, start_index_list2, end_index_list2

modules/llama_cpp_python_hijack.py

@@ -2,6 +2,9 @@
 
 from tqdm import tqdm
 
+from modules import shared
+from modules.cache_utils import process_llamacpp_cache
+
 try:
     import llama_cpp
 except:
@@ -58,6 +61,25 @@ def eval_with_progress(self, tokens: Sequence[int]):
         self.n_tokens += n_tokens
 
 
+def monkey_patch_generate(lib):
+
+    def my_generate(self, *args, **kwargs):
+
+        if shared.args.streaming_llm:
+            new_sequence = args[0]
+            past_sequence = self._input_ids
+
+            # Do the cache trimming for StreamingLLM
+            process_llamacpp_cache(self, new_sequence, past_sequence)
+
+        for output in self.original_generate(*args, **kwargs):
+            yield output
+
+    lib.Llama.original_generate = lib.Llama.generate
+    lib.Llama.generate = my_generate
+
+
 for lib in [llama_cpp, llama_cpp_cuda, llama_cpp_cuda_tensorcores]:
     if lib is not None:
         lib.Llama.eval = eval_with_progress
+        monkey_patch_generate(lib)

modules/loaders.py

@@ -46,6 +46,8 @@
         'no_offload_kqv',
         'row_split',
         'tensorcores',
+        'streaming_llm',
+        'attention_sink_size',
     ],
     'llamacpp_HF': [
         'n_ctx',
@@ -69,6 +71,8 @@
         'no_offload_kqv',
         'row_split',
         'tensorcores',
+        'streaming_llm',
+        'attention_sink_size',
         'llamacpp_HF_info',
     ],
     'ExLlamav2_HF': [

modules/shared.py

@@ -130,6 +130,8 @@
 group.add_argument('--no_offload_kqv', action='store_true', help='Do not offload the  K, Q, V to the GPU. This saves VRAM but reduces the performance.')
 group.add_argument('--cache-capacity', type=str, help='Maximum cache capacity (llama-cpp-python). Examples: 2000MiB, 2GiB. When provided without units, bytes will be assumed.')
 group.add_argument('--row_split', action='store_true', help='Split the model by rows across GPUs. This may improve multi-gpu performance.')
+group.add_argument('--streaming-llm', action='store_true', help='Activates StreamingLLM, which prevents the prompt from ever being reevaluated when old chat messages are removed due to the context length for the model being reached.')
+group.add_argument('--attention-sink-size', type=int, default=5, help='Minimum attention sink length from StreamingLLM.')
 
 # ExLlamaV2
 group = parser.add_argument_group('ExLlamaV2')

modules/text_generation.py

@@ -13,6 +13,7 @@
 from transformers import LogitsProcessorList, is_torch_xpu_available
 
 import modules.shared as shared
+from modules.cache_utils import process_llamacpp_cache
 from modules.callbacks import (
     Iteratorize,
     Stream,
@@ -364,6 +365,12 @@ def generate_reply_HF(question, original_question, seed, state, stopping_strings
         print(decode(input_ids[0], skip_special_tokens=False))
         print()
 
+    # Handle StreamingLLM for llamacpp_HF
+    if shared.model.__class__.__name__ == 'LlamacppHF' and shared.args.streaming_llm:
+        tmp = process_llamacpp_cache(shared.model.model, input_ids[-1].tolist(), shared.model.model._input_ids)
+        shared.model.past_seq = torch.tensor(tmp)
+        shared.model.save_cache()
+
     t0 = time.time()
     try:
         if not is_chat and not shared.is_seq2seq:

modules/ui.py

@@ -97,6 +97,8 @@ def list_model_elements():
         'no_offload_kqv',
         'row_split',
         'tensorcores',
+        'streaming_llm',
+        'attention_sink_size',
         'hqq_backend',
     ]
     if is_torch_xpu_available():

modules/ui_model_menu.py

@@ -117,6 +117,8 @@ def create_ui():
                             shared.gradio['use_flash_attention_2'] = gr.Checkbox(label="use_flash_attention_2", value=shared.args.use_flash_attention_2, info='Set use_flash_attention_2=True while loading the model.')
                             shared.gradio['auto_devices'] = gr.Checkbox(label="auto-devices", value=shared.args.auto_devices)
                             shared.gradio['tensorcores'] = gr.Checkbox(label="tensorcores", value=shared.args.tensorcores, info='NVIDIA only: use llama-cpp-python compiled with tensor cores support. This increases performance on RTX cards.')
+                            shared.gradio['streaming_llm'] = gr.Checkbox(label="streaming_llm", value=shared.args.streaming_llm, info='(experimental) Activate StreamingLLM to avoid re-evaluating the entire prompt when old messages are removed.')
+                            shared.gradio['attention_sink_size'] = gr.Number(label="attention_sink_size", value=shared.args.attention_sink_size)
                             shared.gradio['cpu'] = gr.Checkbox(label="cpu", value=shared.args.cpu, info='llama.cpp: Use llama-cpp-python compiled without GPU acceleration. Transformers: use PyTorch in CPU mode.')
                             shared.gradio['row_split'] = gr.Checkbox(label="row_split", value=shared.args.row_split, info='Split the model by rows across GPUs. This may improve multi-gpu performance.')
                             shared.gradio['no_offload_kqv'] = gr.Checkbox(label="no_offload_kqv", value=shared.args.no_offload_kqv, info='Do not offload the  K, Q, V to the GPU. This saves VRAM but reduces the performance.')