Spaces:
Sleeping
Sleeping
File size: 8,736 Bytes
b786614 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 | """
core.py — The main ProactiveCache user-facing API.
Three lines to make any HuggingFace model O(n):
model = ProactiveCache.apply(model, budget=512)
ProactiveCache.profile(model, tokenizer, corpus="wikitext")
output = model.generate(input_ids, max_new_tokens=500)
"""
from __future__ import annotations
import os
import pickle
import torch
from typing import Optional, Union, List, Dict
from functools import wraps
from .profiler import profile_model as _profile_model
from .prototypes import build_prototypes, save_prototypes, load_prototypes
from .eviction import evict
from .utils import get_device, get_seq_len
class ProactiveCache:
"""
Make any HuggingFace transformer O(n) with proactive KV cache eviction.
Proactive Cache profiles attention head patterns offline (once per model),
then uses the resulting prototype centroids to score token importance at
inference time — without any query lookups. This makes each decode step
O(budget) instead of O(n), achieving up to 3× generation speedup at
2048-token contexts on LLaMA-3.1 8B.
Quickstart:
from proactive_cache import ProactiveCache
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B")
# Apply O(n) eviction (one line)
model = ProactiveCache.apply(model, budget=512)
# Profile once on calibration data (saves to disk)
ProactiveCache.profile(model, tokenizer, corpus="wikitext")
# All inference is now O(n)
output = model.generate(input_ids, max_new_tokens=500)
Licensed under the GNU Affero General Public License v3 (AGPLv3). See LICENSE.
"""
@classmethod
def apply(
cls,
model,
budget: int = 256,
prototype_path: Optional[str] = None,
):
"""
Wrap a HuggingFace model's ``generate()`` with O(n) KV eviction.
After calling ``apply()``, the model's ``generate()`` method will:
1. Run the standard prefill pass.
2. Score token importance using prototypes (or sink+recency if no
prototypes are loaded yet).
3. Prune the KV cache to ``budget`` tokens.
4. Continue auto-regressive decoding against the pruned cache —
each step now attends to B tokens instead of n, achieving O(n)
overall complexity.
Args:
model: Any HuggingFace ``AutoModelForCausalLM`` instance.
budget: Fixed KV cache size kept after eviction (tokens).
prototype_path: Path to prototypes .pkl. If None, will search for
a ``proactive_cache_prototypes.pkl`` in the current
working directory.
Returns:
The same model object, with ``generate()`` patched in-place.
"""
# Resolve prototype path
if prototype_path is None:
prototype_path = "proactive_cache_prototypes.pkl"
prototypes = None
if os.path.exists(prototype_path):
prototypes = load_prototypes(prototype_path)
else:
print(
f"[ProactiveCache] No prototypes at '{prototype_path}'. "
"Using attention-sink + recency fallback. "
"Run ProactiveCache.profile() for best results."
)
# Store config on the model for introspection
model._proactive_config = {
"budget": budget,
"prototype_path": prototype_path,
"active": True,
}
model._proactive_prototypes = prototypes
# Patch generate()
original_generate = model.generate.__func__ if hasattr(model.generate, "__func__") \
else model.generate
@wraps(original_generate)
def proactive_generate(input_ids=None, **kwargs):
return cls._wrapped_generate(
model=model,
original_generate=original_generate,
input_ids=input_ids,
budget=budget,
prototypes=model._proactive_prototypes,
**kwargs,
)
# Bind the patched generate
import types
model.generate = types.MethodType(
lambda self, *args, **kw: proactive_generate(*args, **kw),
model,
)
print(
f"[ProactiveCache] Applied to {type(model).__name__} | "
f"budget={budget} | "
f"prototypes={'loaded' if prototypes else 'none (fallback)'}"
)
return model
@classmethod
def profile(
cls,
model,
tokenizer,
corpus: Union[str, List[str]] = "wikitext",
num_docs: int = 50,
seq_len: int = 512,
n_clusters: int = 4,
save_path: Optional[str] = "proactive_cache_prototypes.pkl",
) -> Dict:
"""
Profile a model's attention patterns and build the prototype library.
Run this once per model. The resulting .pkl file can be shared and
reused for the same model family across different hardware.
Args:
model: HuggingFace CausalLM model (same one passed to apply()).
tokenizer: Corresponding tokenizer.
corpus: Dataset name ("wikitext", "pg19") or list of strings.
num_docs: Number of calibration documents (default 50).
seq_len: Sequence length for profiling (default 512).
n_clusters: KMeans clusters per (layer, head) pair (default 4).
save_path: Where to save the .pkl prototype file.
Returns:
prototypes: The built prototype library dict.
"""
print(f"[ProactiveCache] Starting profiling: corpus={corpus}, "
f"num_docs={num_docs}, seq_len={seq_len}")
patterns = _profile_model(model, tokenizer, corpus, num_docs, seq_len)
prototypes = build_prototypes(patterns, n_clusters=n_clusters, max_seq_len=seq_len)
if save_path:
save_prototypes(prototypes, save_path)
# Update live model if already patched
if hasattr(model, "_proactive_prototypes"):
model._proactive_prototypes = prototypes
print("[ProactiveCache] Live model updated with new prototypes.")
return prototypes
@classmethod
def remove(cls, model):
"""Remove the ProactiveCache patch from a model, restoring original generate()."""
if not hasattr(model, "_proactive_config"):
print("[ProactiveCache] Model is not patched.")
return model
# Restore original generate via re-loading the class method
model.generate = type(model).generate.__get__(model, type(model))
del model._proactive_config
del model._proactive_prototypes
print("[ProactiveCache] Patch removed. Model restored to original generate().")
return model
# ── Internal ──────────────────────────────────────────────────────────────
@staticmethod
def _wrapped_generate(
model,
original_generate,
input_ids,
budget: int,
prototypes: Optional[Dict],
**kwargs,
):
"""
Internal: patched generate() that injects KV eviction after prefill.
Strategy:
1. Run prefill forward pass to build the full KV cache.
2. Evict the cache down to `budget` tokens.
3. Feed only the last input token + pruned cache to generate(),
which continues auto-regressive decoding from that state.
"""
device = get_device(model)
seq_len = input_ids.shape[1]
# If context fits in budget, skip eviction entirely
if seq_len <= budget:
return original_generate(model, input_ids, **kwargs)
# 1. Prefill
with torch.no_grad():
out = model(input_ids, use_cache=True)
past_kv = out.past_key_values
# 2. Evict
past_kv = evict(past_kv, budget, prototypes, seq_len, device)
# 3. Continue generation from pruned KV cache
# Pass only the last token as "prompt" — the KV cache holds the rest
last_token = input_ids[:, -1:]
max_new_tokens = kwargs.pop("max_new_tokens", 100)
return original_generate(
model,
last_token,
past_key_values=past_kv,
max_new_tokens=max_new_tokens,
**kwargs,
)
|