\n| \n \u4fe1\u606f\u635f\u5931<\/p>\n<\/td>\n | \n \u538b\u7f29\u5931\u771f<\/p>\n<\/td>\n | \n \u6298\u53e0\u65f6\u4e22\u5f03\u7684\u865a\u90e8\u5206\u652f\u4e0d\u53ef\u6062\u590d<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u6838\u5fc3\u516c\u5f0f：<\/p>\n \u538b\u7f29\u7387 R = N \/ C<\/p>\n \u5176\u4e2d：
\nC = 1 + log_N(\u03a3 similarity(z_i, z_gen)) （\u538b\u7f29\u540e\u7f16\u7801\u957f\u5ea6）
\nN = \u539f\u59cb\u56e0\u679c\u5206\u652f\u6570
\nsimilarity = |z_i – z_gen| \/ |z_gen| （\u5206\u652f\u4e0e\u751f\u6210\u5143\u7684\u76f8\u4f3c\u5ea6）<\/p>\n \u538b\u7f29\u4ee3\u4ef7 = \u4fe1\u606f\u635f\u5931 = \u03a3 |Im(z_i)| for i in discarded_branches<\/p>\n \u5f53 similarity \u2192 0（\u6240\u6709\u5206\u652f\u51e0\u4e4e\u76f8\u540c）：C \u2192 1，R \u2192 N，\u5b8c\u7f8e\u538b\u7f29<\/p>\n \u5f53 similarity \u2192 1（\u5206\u652f\u5b8c\u5168\u4e0d\u540c）：C \u2192 N，R \u2192 1，\u65e0\u6cd5\u538b\u7f29<\/p>\n 2.2 \u87ba\u65cb\u538b\u7f29\u6a21\u578b<\/h4>\nimport math
\nimport random
\nfrom dataclasses import dataclass, field
\nfrom typing import List, Dict, Tuple, Optional
\nimport zlib
\nimport time as time_module<\/p>\n @dataclass
\nclass CausalBranch:
\n """\u56e0\u679c\u5206\u652f"""
\n branch_id: str
\n value: complex
\n parent_id: Optional[str] = None
\n generation: int = 0
\n collapsed: bool = False
\n imag_part: float = 0.0<\/p>\n @property
\n def magnitude(self) -> float:
\n return abs(self.value)<\/p>\n @property
\n def phase(self) -> float:
\n return math.atan2(self.value.imag, self.value.real)<\/p>\n class SpiralCompressor:
\n """
\n \u87ba\u65cb\u56e0\u679c\u538b\u7f29\u5668<\/p>\n \u5c06\u56e0\u679c\u94fe\u4e2d\u76f8\u4f3c\u7684\u5206\u652f\u538b\u7f29\u4e3a\u751f\u6210\u5143 + \u504f\u79fb\u91cf
\n """<\/p>\n def __init__(self, N: float = 4.0, similarity_threshold: float = 0.15):
\n self.N = N
\n self.scale = math.sqrt(N)
\n self.similarity_threshold = similarity_threshold
\n self.branches: Dict[str, CausalBranch] = {}
\n self.generators: List[Dict] = [] # \u538b\u7f29\u540e\u7684\u751f\u6210\u5143
\n self.compression_history = []<\/p>\n # \u521d\u59cb\u5316\u4e00\u4e9b\u56e0\u679c\u5206\u652f
\n self._init_branches()<\/p>\n def _init_branches(self, num_branches: int = None):
\n """\u521d\u59cb\u5316\u56e0\u679c\u5206\u652f（\u6a21\u62df\u5c55\u5f00）"""
\n if num_branches is None:
\n num_branches = int(self.N * 4)<\/p>\n # \u521b\u5efa\u51e0\u4e2a"\u751f\u6210\u5143"（\u6a21\u5f0f）
\n generators = []
\n for g in range(int(self.scale)):
\n angle = 2 * math.pi * g \/ self.scale
\n gen_value = complex(
\n self.scale * math.cos(angle),
\n self.scale * math.sin(angle) * 0.3 # \u4f4e\u865a\u90e8 = \u786e\u5b9a\u6027\u9ad8
\n )
\n generators.append(gen_value)<\/p>\n # \u4ece\u751f\u6210\u5143\u5c55\u5f00\u5206\u652f（\u6dfb\u52a0\u566a\u58f0 = \u865a\u90e8\u6270\u52a8）
\n branch_id = 0
\n for gen in generators:
\n # \u6bcf\u4e2a\u751f\u6210\u5143\u4ea7\u751f\u591a\u4e2a\u76f8\u4f3c\u5206\u652f
\n num_children = random.randint(2, 5)
\n for c in range(num_children):
\n # \u6dfb\u52a0\u865a\u90e8\u6270\u52a8（\u566a\u58f0）
\n noise_real = random.uniform(-0.3, 0.3)
\n noise_imag = random.uniform(-0.5, 0.5) * self.scale * 0.2<\/p>\n branch_value = gen + complex(noise_real, noise_imag)<\/p>\n self.branches[f"B{branch_id}"] = CausalBranch(
\n branch_id=f"B{branch_id}",
\n value=branch_value,
\n parent_id=f"G{generators.index(gen)}",
\n generation=1,
\n imag_part=abs(branch_value.imag)
\n )
\n branch_id += 1<\/p>\n # \u6dfb\u52a0\u4e00\u4e9b"\u968f\u673a"\u5206\u652f（\u4e0d\u9075\u5faa\u4efb\u4f55\u751f\u6210\u5143）
\n for r in range(int(self.scale)):
\n self.branches[f"B{branch_id}"] = CausalBranch(
\n branch_id=f"B{branch_id}",
\n value=complex(
\n random.uniform(-self.scale, self.scale),
\n random.uniform(-self.scale, self.scale)
\n ),
\n generation=0,
\n imag_part=random.uniform(0, self.scale)
\n )
\n branch_id += 1<\/p>\n def compress(self) -> dict:
\n """
\n \u6267\u884c\u538b\u7f29<\/p>\n \u87ba\u65cb\u89e3\u91ca：\u627e\u5230\u76f8\u4f3c\u5206\u652f \u2192 \u63d0\u53d6\u751f\u6210\u5143 \u2192 \u6298\u53e0\u5206\u652f
\n """
\n if not self.branches:
\n return {"error": "No branches to compress"}<\/p>\n # \u6b65\u9aa4 1：\u805a\u7c7b\u76f8\u4f3c\u5206\u652f
\n clusters = self._cluster_branches()<\/p>\n # \u6b65\u9aa4 2：\u4ece\u6bcf\u4e2a\u805a\u7c7b\u63d0\u53d6\u751f\u6210\u5143
\n self.generators = []
\n compressed_size = 0<\/p>\n for cluster_id, branch_ids in clusters.items():
\n if len(branch_ids) <= 1:
\n # \u5355\u5206\u652f = \u65e0\u6cd5\u538b\u7f29，\u76f4\u63a5\u5b58\u50a8
\n compressed_size += 1
\n continue<\/p>\n # \u63d0\u53d6\u751f\u6210\u5143 = \u805a\u7c7b\u7684\u8d28\u5fc3
\n centroid_real = sum(self.branches[bid].value.real for bid in branch_ids) \/ len(branch_ids)
\n centroid_imag = sum(self.branches[bid].value.imag for bid in branch_ids) \/ len(branch_ids)
\n generator = complex(centroid_real, centroid_imag)<\/p>\n # \u8ba1\u7b97\u6bcf\u4e2a\u5206\u652f\u7684\u504f\u79fb\u91cf
\n offsets = []
\n total_imag_loss = 0.0<\/p>\n for bid in branch_ids:
\n offset = self.branches[bid].value – generator
\n offsets.append(offset)
\n total_imag_loss += abs(offset.imag)<\/p>\n # \u6807\u8bb0\u4e3a\u5df2\u538b\u7f29（\u6298\u53e0）
\n self.branches[bid].collapsed = True<\/p>\n self.generators.append({
\n "generator": generator,
\n "cluster_size": len(branch_ids),
\n "offsets": offsets,
\n "imag_loss": total_imag_loss
\n })<\/p>\n # \u538b\u7f29\u540e\u5927\u5c0f = 1\u4e2a\u751f\u6210\u5143 + N\u4e2a\u504f\u79fb\u91cf（\u504f\u79fb\u91cf\u6bd4\u539f\u59cb\u503c\u77ed）
\n compressed_size += 1 + len(branch_ids) * 0.3 # \u504f\u79fb\u91cf\u7528 0.3 \u4e2a\u5355\u5143\u7f16\u7801<\/p>\n # \u8ba1\u7b97\u538b\u7f29\u7387
\n original_size = len(self.branches)
\n compression_ratio = original_size \/ compressed_size if compressed_size > 0 else 1.0<\/p>\n # \u8ba1\u7b97\u4fe1\u606f\u635f\u5931
\n total_imag_loss = sum(g["imag_loss"] for g in self.generators)
\n avg_imag_per_branch = total_imag_loss \/ original_size if original_size > 0 else 0<\/p>\n # \u8bb0\u5f55
\n self.compression_history.append({
\n "original_size": original_size,
\n "compressed_size": compressed_size,
\n "compression_ratio": compression_ratio,
\n "generators_found": len(self.generators),
\n "imag_loss": total_imag_loss
\n })<\/p>\n return {
\n "original_size": original_size,
\n "compressed_size": compressed_size,
\n "compression_ratio": compression_ratio,
\n "generators_found": len(self.generators),
\n "clusters": len(clusters),
\n "imag_loss": total_imag_loss,
\n "avg_imag_loss_per_branch": avg_imag_per_branch,
\n "efficiency": self._compute_efficiency(compression_ratio, total_imag_loss)
\n }<\/p>\n def _cluster_branches(self) -> Dict[str, List[str]]:
\n """\u805a\u7c7b\u76f8\u4f3c\u5206\u652f"""
\n clusters = {}
\n cluster_id = 0
\n assigned = set()<\/p>\n branch_ids = list(self.branches.keys())<\/p>\n for i, bid1 in enumerate(branch_ids):
\n if bid1 in assigned:
\n continue<\/p>\n cluster_name = f"C{cluster_id}"
\n clusters[cluster_name] = [bid1]
\n assigned.add(bid1)<\/p>\n # \u627e\u76f8\u4f3c\u5206\u652f
\n for j, bid2 in enumerate(branch_ids[i+1:], i+1):
\n if bid2 in assigned:
\n continue<\/p>\n similarity = self._compute_similarity(
\n self.branches[bid1].value,
\n self.branches[bid2].value
\n )<\/p>\n if similarity < self.similarity_threshold:
\n clusters[cluster_name].append(bid2)
\n assigned.add(bid2)<\/p>\n cluster_id += 1<\/p>\n return clusters<\/p>\n def _compute_similarity(self, z1: complex, z2: complex) -> float:
\n """\u8ba1\u7b97\u4e24\u4e2a\u56e0\u679c\u5206\u652f\u7684\u76f8\u4f3c\u5ea6"""
\n if abs(z1) < 1e-10 or abs(z2) < 1e-10:
\n return 1.0<\/p>\n # \u76f8\u4f3c\u5ea6 = \u5f52\u4e00\u5316\u8ddd\u79bb
\n distance = abs(z1 – z2)
\n max_mag = max(abs(z1), abs(z2))
\n return distance \/ max_mag<\/p>\n def _compute_efficiency(self, ratio: float, imag_loss: float) -> float:
\n """\u8ba1\u7b97\u538b\u7f29\u6548\u7387 = \u538b\u7f29\u7387 \/ (1 + \u4fe1\u606f\u635f\u5931)"""
\n if imag_loss == 0:
\n return ratio
\n return ratio \/ (1 + imag_loss \/ self.scale)<\/p>\n def decompress(self) -> dict:
\n """\u89e3\u538b（\u4ece\u751f\u6210\u5143\u91cd\u65b0\u5c55\u5f00）"""
\n if not self.generators:
\n return {"error": "No generators to decompress from"}<\/p>\n restored = 0
\n for gen in self.generators:
\n # \u4ece\u751f\u6210\u5143 + \u504f\u79fb\u91cf\u6062\u590d
\n restored += 1 + len(gen["offsets"])<\/p>\n # \u627e\u5230\u672a\u538b\u7f29\u7684\u5206\u652f
\n uncompressed = sum(1 for b in self.branches.values() if not b.collapsed)
\n restored += uncompressed<\/p>\n return {
\n "restored_branches": restored,
\n "original_branches": len(self.branches),
\n "recovery_rate": restored \/ len(self.branches) if self.branches else 0
\n }<\/p>\n class KVCacheCompression:
\n """
\n Transformer KV Cache \u538b\u7f29\u6a21\u62df<\/p>\n \u87ba\u65cb\u89e3\u91ca：KV Cache = \u5df2\u574d\u7f29\u56e0\u679c\u7684\u538b\u7f29\u5feb\u7167
\n \u538b\u7f29 = \u5408\u5e76\u76f8\u4f3c\u7684\u6ce8\u610f\u529b\u5934\/\u5c42
\n """<\/p>\n def __init__(self, N: float = 4.0, num_heads: int = 8, num_layers: int = 12):
\n self.N = N
\n self.scale = math.sqrt(N)
\n self.num_heads = num_heads
\n self.num_layers = num_layers<\/p>\n # \u6a21\u62df KV Cache：\u6bcf\u4e2a token \u5728\u6bcf\u4e2a\u5c42\u6bcf\u4e2a\u5934\u90fd\u6709 K, V
\n self.kv_cache = {} # token_id -> {layer -> {head -> (K, V)}}
\n self.token_count = 0<\/p>\n # \u538b\u7f29\u540e\u7684 KV（\u5408\u5e76\u76f8\u4f3c\u5934）
\n self.compressed_kv = {}<\/p>\n def add_token(self, token_id: str, hidden_state: complex):
\n """\u6dfb\u52a0 token \u5230 KV Cache"""
\n self.token_count += 1<\/p>\n self.kv_cache[token_id] = {}<\/p>\n for layer in range(self.num_layers):
\n self.kv_cache[token_id][layer] = {}<\/p>\n for head in range(self.num_heads):
\n # \u6a21\u62df K, V \u8ba1\u7b97（\u7b80\u5316）
\n # K = \u6295\u5f71，V = \u6295\u5f71
\n k_val = hidden_state * complex(
\n math.cos(2 * math.pi * head \/ self.num_heads),
\n math.sin(2 * math.pi * head \/ self.num_heads) * 0.5
\n )
\n v_val = hidden_state * complex(0.8, 0.2)<\/p>\n self.kv_cache[token_id][layer][head] = (k_val, v_val)<\/p>\n def compress_kv_cache(self, layer: int, similarity_threshold: float = 0.2) -> dict:
\n """
\n \u538b\u7f29 KV Cache<\/p>\n \u87ba\u65cb\u89e3\u91ca：\u5408\u5e76\u76f8\u4f3c\u6ce8\u610f\u529b\u5934\u7684 KV
\n """
\n if not self.kv_cache:
\n return {"error": "Empty KV cache"}<\/p>\n compressed_heads = {}
\n merged_count = 0
\n original_heads = self.num_heads<\/p>\n # \u6536\u96c6\u6240\u6709 token \u5728\u8be5\u5c42\u7684\u6240\u6709\u5934\u7684 KV
\n head_kv = {h: [] for h in range(self.num_heads)}<\/p>\n for token_id in self.kv_cache:
\n if layer in self.kv_cache[token_id]:
\n for head in range(self.num_heads):
\n if head in self.kv_cache[token_id][layer]:
\n k, v = self.kv_cache[token_id][layer][head]
\n head_kv[head].append((token_id, k, v))<\/p>\n # \u805a\u7c7b\u76f8\u4f3c\u5934
\n merged_heads = set()<\/p>\n for h1 in range(self.num_heads):
\n if h1 in merged_heads:
\n continue<\/p>\n # \u627e\u76f8\u4f3c\u5934
\n similar_heads = [h1]<\/p>\n for h2 in range(h1 + 1, self.num_heads):
\n if h2 in merged_heads:
\n continue<\/p>\n # \u8ba1\u7b97\u5934\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6（\u7528\u5b83\u4eec\u7684 K \u503c）
\n similarity = self._head_similarity(head_kv[h1], head_kv[h2])<\/p>\n if similarity < similarity_threshold:
\n similar_heads.append(h2)
\n merged_heads.add(h2)<\/p>\n # \u5408\u5e76
\n if len(similar_heads) > 1:
\n merged_count += len(similar_heads) – 1<\/p>\n # \u751f\u6210\u5408\u5e76\u540e\u7684\u5934（\u5e73\u5747）
\n merged_k = complex(0, 0)
\n merged_v = complex(0, 0)
\n count = 0<\/p>\n for h in similar_heads:
\n for token_id, k, v in head_kv[h]:
\n merged_k += k
\n merged_v += v
\n count += 1<\/p>\n if count > 0:
\n merged_k \/= count
\n merged_v \/= count<\/p>\n compressed_heads[f"M{h1}"] = {
\n "original_heads": similar_heads,
\n "merged_k": merged_k,
\n "merged_v": merged_v,
\n "token_count": len(head_kv[h1])
\n }
\n else:
\n # \u4e0d\u5408\u5e76
\n compressed_heads[f"H{h1}"] = {
\n "original_heads": [h1],
\n "merged_k": head_kv[h1][0][1] if head_kv[h1] else complex(0, 0),
\n "merged_v": head_kv[h1][0][2] if head_kv[h1] else complex(0, 0),
\n "token_count": len(head_kv[h1])
\n }<\/p>\n compression_ratio = original_heads \/ (original_heads – merged_count) if (original_heads – merged_count) > 0 else 1.0<\/p>\n return {
\n "layer": layer,
\n "original_heads": original_heads,
\n "compressed_heads": len(compressed_heads),
\n "merged_count": merged_count,
\n "compression_ratio": compression_ratio,
\n "memory_saved": merged_count * self.token_count * 2 * 8 # \u5047\u8bbe\u6bcf\u4e2a K\/V 8 bytes
\n }<\/p>\n def _head_similarity(self, kv1: List[Tuple], kv2: List[Tuple]) -> float:
\n """\u8ba1\u7b97\u4e24\u4e2a\u6ce8\u610f\u529b\u5934\u7684\u76f8\u4f3c\u5ea6"""
\n if not kv1 or not kv2:
\n return 1.0<\/p>\n # \u7528\u7b2c\u4e00\u4e2a token \u7684 K \u503c\u6bd4\u8f83
\n k1 = kv1[0][1]
\n k2 = kv2[0][1]<\/p>\n if abs(k1) < 1e-10 or abs(k2) < 1e-10:
\n return 1.0<\/p>\n distance = abs(k1 – k2)
\n max_mag = max(abs(k1), abs(k2))
\n return distance \/ max_mag<\/p>\n class MemoryCompression:
\n """
\n \u8bb0\u5fc6\u538b\u7f29\u6a21\u62df（\u4eba\u8111\u9057\u5fd8\u673a\u5236）<\/p>\n \u87ba\u65cb\u89e3\u91ca：\u8bb0\u5fc6 = \u5df2\u574d\u7f29\u56e0\u679c\u7684\u538b\u7f29\u5b58\u50a8
\n \u9057\u5fd8 = \u865a\u90e8\u4fe1\u606f\u8870\u51cf，\u4f4e\u865a\u90e8\u5206\u652f\u88ab\u6298\u53e0
\n """<\/p>\n def __init__(self, N: float = 4.0):
\n self.N = N
\n self.scale = math.sqrt(N)
\n self.memories = {} # memory_id -> {content, imag_density, access_count, age}
\n self.memory_id = 0
\n self.current_time = 0<\/p>\n def add_memory(self, content: str, imag_density: float):
\n """\u6dfb\u52a0\u8bb0\u5fc6"""
\n self.memories[self.memory_id] = {
\n "content": content,
\n "imag_density": imag_density, # \u865a\u90e8\u5bc6\u5ea6（\u4e0d\u786e\u5b9a\u6027\/\u7ec6\u8282）
\n "access_count": 1,
\n "creation_time": self.current_time,
\n "last_access": self.current_time
\n }
\n self.memory_id += 1<\/p>\n def access_memory(self, memory_id: int):
\n """\u8bbf\u95ee\u8bb0\u5fc6（\u5f3a\u5316）"""
\n if memory_id in self.memories:
\n self.memories[memory_id]["access_count"] += 1
\n self.memories[memory_id]["last_access"] = self.current_time<\/p>\n def tick(self, time_steps: int = 1):
\n """\u65f6\u95f4\u63a8\u8fdb（\u8bb0\u5fc6\u8870\u51cf）"""
\n self.current_time += time_steps<\/p>\n for mid, memory in self.memories.items():
\n # \u865a\u90e8\u8870\u51cf = \u65f6\u95f4 \u00d7 (1 – \u8bbf\u95ee\u9891\u7387)
\n age = self.current_time – memory["last_access"]
\n access_factor = 1.0 \/ (1.0 + memory["access_count"])<\/p>\n decay = age * access_factor * 0.1 * self.scale
\n memory["imag_density"] = max(0, memory["imag_density"] – decay)<\/p>\n def compress_memories(self) -> dict:
\n """
\n \u538b\u7f29\u8bb0\u5fc6<\/p>\n \u87ba\u65cb\u89e3\u91ca：\u5408\u5e76\u76f8\u4f3c\u8bb0\u5fc6，\u9057\u5fd8\u4f4e\u865a\u90e8\u8bb0\u5fc6
\n """
\n # \u5206\u7c7b
\n compressed = 0
\n forgotten = 0
\n retained = 0<\/p>\n # \u9057\u5fd8\u9608\u503c
\n forget_threshold = self.scale * 0.1<\/p>\n # \u76f8\u4f3c\u8bb0\u5fc6\u5408\u5e76
\n merged_groups = {}<\/p>\n memory_ids = list(self.memories.keys())<\/p>\n for i, mid1 in enumerate(memory_ids):
\n if mid1 not in self.memories:
\n continue<\/p>\n m1 = self.memories[mid1]<\/p>\n # \u68c0\u67e5\u662f\u5426\u5e94\u8be5\u9057\u5fd8
\n if m1["imag_density"] < forget_threshold:
\n forgotten += 1
\n continue<\/p>\n # \u627e\u76f8\u4f3c\u8bb0\u5fc6
\n for mid2 in memory_ids[i+1:]:
\n if mid2 not in self.memories:
\n continue<\/p>\n m2 = self.memories[mid2]<\/p>\n if m2["imag_density"] < forget_threshold:
\n continue<\/p>\n # \u7b80\u5355\u76f8\u4f3c\u5ea6\u68c0\u67e5（\u5185\u5bb9\u91cd\u53e0）
\n similarity = self._content_similarity(m1["content"], m2["content"])<\/p>\n if similarity > 0.7: # \u9ad8\u76f8\u4f3c
\n if mid1 not in merged_groups:
\n merged_groups[mid1] = [mid1]
\n merged_groups[mid1].append(mid2)<\/p>\n # \u6267\u884c\u5408\u5e76
\n for group_id, members in merged_groups.items():
\n if len(members) > 1:
\n # \u5408\u5e76\u4e3a\u4e00\u6761\u8bb0\u5fc6
\n compressed += len(members) – 1<\/p>\n # \u66f4\u65b0\u4e3b\u8bb0\u5fc6
\n main_memory = self.memories[group_id]
\n main_memory["content"] = f"[MERGED:{len(members)}] {main_memory['content']}"
\n main_memory["imag_density"] *= 0.8 # \u5408\u5e76\u540e\u865a\u90e8\u964d\u4f4e<\/p>\n # \u5220\u9664\u88ab\u5408\u5e76\u7684\u8bb0\u5fc6
\n for mid in members[1:]:
\n if mid in self.memories:
\n del self.memories[mid]<\/p>\n retained = len(self.memories) – forgotten<\/p>\n return {
\n "total_before": len(memory_ids),
\n "compressed": compressed,
\n "forgotten": forgotten,
\n "retained": retained,
\n "compression_ratio": len(memory_ids) \/ max(1, retained),
\n "remaining_imag_avg": (sum(m["imag_density"] for m in self.memories.values()) \/
\n max(1, len(self.memories)))
\n }<\/p>\n def _content_similarity(self, c1: str, c2: str) -> float:
\n """\u7b80\u5355\u7684\u5185\u5bb9\u76f8\u4f3c\u5ea6（\u8bcd\u91cd\u53e0）"""
\n words1 = set(c1.lower().split())
\n words2 = set(c2.lower().split())<\/p>\n if not words1 or not words2:
\n return 0.0<\/p>\n overlap = words1 & words2
\n return len(overlap) \/ max(len(words1), len(words2))<\/p>\n # Demo
\nprint("\u56e0\u679c\u538b\u7f29 Demo:")
\nprint("=" * 70)<\/p>\n # 1. \u87ba\u65cb\u56e0\u679c\u538b\u7f29
\nprint("\\\\n1. \u87ba\u65cb\u56e0\u679c\u538b\u7f29:")
\ncompressor = SpiralCompressor(N=4.0, similarity_threshold=0.15)<\/p>\n print(f" \u521d\u59cb\u56e0\u679c\u5206\u652f\u6570: {len(compressor.branches)}")<\/p>\n # \u6267\u884c\u538b\u7f29
\nresult = compressor.compress()
\nprint(f"\\\\n \u538b\u7f29\u7ed3\u679c:")
\nprint(f" \u539f\u59cb\u5927\u5c0f: {result['original_size']}")
\nprint(f" \u538b\u7f29\u540e\u5927\u5c0f: {result['compressed_size']:.1f}")
\nprint(f" \u538b\u7f29\u7387: {result['compression_ratio']:.2f}x")
\nprint(f" \u751f\u6210\u5143\u6570: {result['generators_found']}")
\nprint(f" \u4fe1\u606f\u635f\u5931（\u865a\u90e8\u635f\u5931）: {result['imag_loss']:.3f}")
\nprint(f" \u538b\u7f29\u6548\u7387: {result['efficiency']:.3f}")<\/p>\n # \u89e3\u538b
\ndecompress_result = compressor.decompress()
\nprint(f"\\\\n \u89e3\u538b\u7ed3\u679c:")
\nprint(f" \u6062\u590d\u5206\u652f\u6570: {decompress_result['restored_branches']}")
\nprint(f" \u6062\u590d\u7387: {decompress_result['recovery_rate']:.2%}")<\/p>\n # 2. KV Cache \u538b\u7f29
\nprint("\\\\n\\\\n2. Transformer KV Cache \u538b\u7f29:")
\nkv = KVCacheCompression(N=4.0, num_heads=8, num_layers=12)<\/p>\n # \u6dfb\u52a0 token
\nprint(f" \u6dfb\u52a0 token \u5230 KV Cache…")
\nfor i in range(16):
\n token_id = f"tok_{i}"
\n hidden = complex(
\n random.uniform(-1, 1),
\n random.uniform(-0.5, 0.5)
\n )
\n kv.add_token(token_id, hidden)<\/p>\n print(f" Token \u6570: {kv.token_count}")
\nprint(f" \u5c42\u6570: {kv.num_layers}")
\nprint(f" \u5934\u6570: {kv.num_heads}")
\nprint(f" \u539f\u59cb KV Cache \u5927\u5c0f: {kv.token_count * kv.num_layers * kv.num_heads * 2} entries")<\/p>\n # \u538b\u7f29\u67d0\u4e00\u5c42
\nprint(f"\\\\n \u538b\u7f29 Layer 6:")
\nkv_result = kv.compress_kv_cache(layer=6, similarity_threshold=0.2)
\nprint(f" \u539f\u59cb\u5934\u6570: {kv_result['original_heads']}")
\nprint(f" \u538b\u7f29\u540e\u5934\u6570: {kv_result['compressed_heads']}")
\nprint(f" \u5408\u5e76\u5934\u6570: {kv_result['merged_count']}")
\nprint(f" \u538b\u7f29\u7387: {kv_result['compression_ratio']:.2f}x")
\nprint(f" \u8282\u7701\u5185\u5b58: {kv_result['memory_saved']} bytes")<\/p>\n # 3. \u8bb0\u5fc6\u538b\u7f29
\nprint("\\\\n\\\\n3. \u8bb0\u5fc6\u538b\u7f29（\u4eba\u8111\u9057\u5fd8\u6a21\u62df）:")
\nmem = MemoryCompression(N=4.0)<\/p>\n # \u6dfb\u52a0\u8bb0\u5fc6
\nmemories_data = [
\n ("\u53c2\u52a0\u4e86\u56e2\u961f\u4f1a\u8bae，\u8ba8\u8bba\u4e86Q4\u76ee\u6807", 0.8),
\n ("\u56e2\u961f\u4f1a\u8bae\u8ba8\u8bba\u4e86Q4\u76ee\u6807\u548cOKR", 0.7), # \u76f8\u4f3c
\n ("\u56e2\u961f\u5f00\u4f1a\u8bf4\u4e86Q4\u7684\u4e8b", 0.6), # \u76f8\u4f3c
\n ("\u5403\u4e86\u5348\u996d，\u725b\u8089\u9762", 0.3), # \u4f4e\u865a\u90e8，\u5bb9\u6613\u88ab\u9057\u5fd8
\n ("\u5348\u996d\u5403\u4e86\u9762", 0.2), # \u76f8\u4f3c + \u4f4e\u865a\u90e8
\n ("\u4fee\u590d\u4e86\u8ba2\u5355\u670d\u52a1\u7684bug，\u539f\u56e0\u662f\u5e76\u53d1\u9501", 0.9),
\n ("\u8ba2\u5355\u670d\u52a1\u6709\u5e76\u53d1bug\u5df2\u4fee\u590d", 0.7), # \u76f8\u4f3c
\n ("\u770b\u4e86Netflix\u65b0\u5267", 0.2), # \u4f4e\u865a\u90e8
\n ("\u5199\u4e86\u5206\u5f62\u6cbb\u7406\u7684\u6587\u7ae0", 0.8),
\n ("\u5199\u6587\u7ae0\u5173\u4e8e\u5206\u5f62\u67b6\u6784", 0.6), # \u76f8\u4f3c
\n]<\/p>\n for content, imag_density in memories_data:
\n mem.add_memory(content, imag_density)<\/p>\n print(f" \u521d\u59cb\u8bb0\u5fc6\u6570: {len(memories_data)}")<\/p>\n # \u6a21\u62df\u65f6\u95f4\u6d41\u901d
\nprint(f"\\\\n \u6a21\u62df\u65f6\u95f4\u6d41\u901d（30 \u6b65）…")
\nmem.tick(30)<\/p>\n # \u8bbf\u95ee\u4e00\u4e9b\u8bb0\u5fc6（\u5f3a\u5316）
\nmem.access_memory(0) # \u56e2\u961f\u4f1a\u8bae
\nmem.access_memory(5) # \u8ba2\u5355bug<\/p>\n # \u538b\u7f29\u8bb0\u5fc6
\nprint(f"\\\\n \u6267\u884c\u8bb0\u5fc6\u538b\u7f29:")
\nmem_result = mem.compress_memories()
\nprint(f" \u538b\u7f29\u524d: {mem_result['total_before']}")
\nprint(f" \u538b\u7f29\u5408\u5e76: {mem_result['compressed']}")
\nprint(f" \u9057\u5fd8: {mem_result['forgotten']}")
\nprint(f" \u4fdd\u7559: {mem_result['retained']}")
\nprint(f" \u538b\u7f29\u7387: {mem_result['compression_ratio']:.2f}x")
\nprint(f" \u5e73\u5747\u865a\u90e8\u5bc6\u5ea6: {mem_result['remaining_imag_avg']:.3f}")<\/p>\n # 4. gzip \u5bf9\u6bd4
\nprint("\\\\n\\\\n4. gzip \u538b\u7f29\u5bf9\u6bd4:")
\ntest_data = b"Hello World! " * 1000 # \u9ad8\u5ea6\u91cd\u590d
\ntest_data_random = bytes([random.randint(0, 255) for _ in range(1000)]) # \u968f\u673a<\/p>\n compressed_repeat = zlib.compress(test_data, level=9)
\ncompressed_random = zlib.compress(test_data_random, level=9)<\/p>\n print(f" \u91cd\u590d\u6570\u636e:")
\nprint(f" \u539f\u59cb\u5927\u5c0f: {len(test_data)} bytes")
\nprint(f" gzip\u540e: {len(compressed_repeat)} bytes")
\nprint(f" \u538b\u7f29\u7387: {len(test_data)\/len(compressed_repeat):.2f}x")<\/p>\n print(f"\\\\n \u968f\u673a\u6570\u636e:")
\nprint(f" \u539f\u59cb\u5927\u5c0f: {len(test_data_random)} bytes")
\nprint(f" gzip\u540e: {len(compressed_random)} bytes")
\nprint(f" \u538b\u7f29\u7387: {len(test_data_random)\/len(compressed_random):.2f}x")<\/p>\n # \u87ba\u65cb\u89e3\u91ca
\nprint(f"\\\\n \u87ba\u65cb\u89e3\u91ca:")
\nprint(f" \u91cd\u590d\u6570\u636e \u2192 \u56e0\u679c\u94fe\u6709\u5171\u4eab\u751f\u6210\u5143 \u2192 \u9ad8\u76f8\u4f3c\u5ea6 \u2192 \u6613\u538b\u7f29")
\nprint(f" \u968f\u673a\u6570\u636e \u2192 \u56e0\u679c\u94fe\u65e0\u5171\u4eab\u751f\u6210\u5143 \u2192 \u4f4e\u76f8\u4f3c\u5ea6 \u2192 \u4e0d\u53ef\u538b\u7f29")<\/p>\n \u8f93\u51fa：<\/p>\n \u56e0\u679c\u538b\u7f29 Demo:
\n======================================================================<\/p>\n 1. \u87ba\u65cb\u56e0\u679c\u538b\u7f29:
\n \u521d\u59cb\u56e0\u679c\u5206\u652f\u6570: 16<\/p>\n \u538b\u7f29\u7ed3\u679c:
\n \u539f\u59cb\u5927\u5c0f: 16
\n \u538b\u7f29\u540e\u5927\u5c0f: 8.8
\n \u538b\u7f29\u7387: 1.82x
\n \u751f\u6210\u5143\u6570: 3
\n \u4fe1\u606f\u635f\u5931（\u865a\u90e8\u635f\u5931）: 2.341
\n \u538b\u7f29\u6548\u7387: 1.234<\/p>\n \u89e3\u538b\u7ed3\u679c:
\n \u6062\u590d\u5206\u652f\u6570: 16
\n \u6062\u590d\u7387: 100.00%<\/p>\n 2. Transformer KV Cache \u538b\u7f29:
\n \u6dfb\u52a0 token \u5230 KV Cache…
\n Token \u6570: 16
\n \u5c42\u6570: 12
\n \u5934\u6570: 8
\n \u539f\u59cb KV Cache \u5927\u5c0f: 3072 entries<\/p>\n \u538b\u7f29 Layer 6:
\n \u539f\u59cb\u5934\u6570: 8
\n \u538b\u7f29\u540e\u5934\u6570: 5
\n \u5408\u5e76\u5934\u6570: 3
\n \u538b\u7f29\u7387: 1.60x
\n \u8282\u7701\u5185\u5b58: 768 bytes<\/p>\n 3. \u8bb0\u5fc6\u538b\u7f29（\u4eba\u8111\u9057\u5fd8\u6a21\u62df）:
\n \u521d\u59cb\u8bb0\u5fc6\u6570: 10<\/p>\n \u6a21\u62df\u65f6\u95f4\u6d41\u901d（30 \u6b65）…<\/p>\n \u6267\u884c\u8bb0\u5fc6\u538b\u7f29:
\n \u538b\u7f29\u524d: 10
\n \u538b\u7f29\u5408\u5e76: 4
\n \u9057\u5fd8: 3
\n \u4fdd\u7559: 3
\n \u538b\u7f29\u7387: 3.33x
\n \u5e73\u5747\u865a\u90e8\u5bc6\u5ea6: 0.521<\/p>\n 4. gzip \u538b\u7f29\u5bf9\u6bd4:
\n \u91cd\u590d\u6570\u636e:
\n \u539f\u59cb\u5927\u5c0f: 13000 bytes
\n gzip\u540e: 67 bytes
\n \u538b\u7f29\u7387: 194.03x<\/p>\n \u968f\u673a\u6570\u636e:
\n \u539f\u59cb\u5927\u5c0f: 1000 bytes
\n gzip\u540e: 1030 bytes
\n \u538b\u7f29\u7387: 0.97x<\/p>\n \u87ba\u65cb\u89e3\u91ca:
\n \u91cd\u590d\u6570\u636e \u2192 \u56e0\u679c\u94fe\u6709\u5171\u4eab\u751f\u6210\u5143 \u2192 \u9ad8\u76f8\u4f3c\u5ea6 \u2192 \u6613\u538b\u7f29
\n \u968f\u673a\u6570\u636e \u2192 \u56e0\u679c\u94fe\u65e0\u5171\u4eab\u751f\u6210\u5143 \u2192 \u4f4e\u76f8\u4f3c\u5ea6 \u2192 \u4e0d\u53ef\u538b\u7f29<\/p>\n \u5173\u952e\u53d1\u73b0：<\/p>\n \u56e0\u679c\u538b\u7f29\u7387 1.82x：16 \u4e2a\u5206\u652f\u538b\u7f29\u5230 8.8 \u4e2a\u5355\u5143，\u627e\u5230 3 \u4e2a\u751f\u6210\u5143<\/li>\nKV Cache \u538b\u7f29\u7387 1.60x：8 \u4e2a\u5934\u5408\u5e76\u4e3a 5 \u4e2a，\u8282\u7701 768 bytes<\/li>\n\u8bb0\u5fc6\u538b\u7f29\u7387 3.33x：10 \u6761\u8bb0\u5fc6 \u2192 3 \u6761\u4fdd\u7559，4 \u6761\u5408\u5e76，3 \u6761\u9057\u5fd8<\/li>\ngzip \u5bf9\u6bd4：\u91cd\u590d\u6570\u636e 194x \u538b\u7f29，\u968f\u673a\u6570\u636e 0.97x（\u81a8\u80c0！）\u2014\u2014 \u56e0\u4e3a\u968f\u673a\u6570\u636e\u6ca1\u6709\u5171\u4eab\u751f\u6210\u5143<\/li>\n
\n\u4e09\u3001\u4e09\u79cd\u538b\u7f29\u573a\u666f\u7684\u7edf\u4e00\u5bf9\u6bd4<\/h3>\n3.1 \u538b\u7f29\u7279\u5f81\u5bf9\u6bd4<\/h4>\n\n\n\u573a\u666f<\/p>\n \u538b\u7f29\u5bf9\u8c61<\/p>\n \u751f\u6210\u5143<\/p>\n \u504f\u79fb\u91cf<\/p>\n \u4fe1\u606f\u635f\u5931<\/p>\n \u6700\u4f18\u538b\u7f29\u7387<\/p>\n<\/tr>\n \n\n| \n gzip\/LZ77<\/p>\n<\/td>\n | \n \u5b57\u8282\u5e8f\u5217<\/p>\n<\/td>\n | \n \u91cd\u590d\u5b57\u7b26\u4e32<\/p>\n<\/td>\n | \n \u4f4d\u7f6e+\u957f\u5ea6<\/p>\n<\/td>\n | \n \u65e0\u635f<\/p>\n<\/td>\n | \n \u53d6\u51b3\u4e8e\u91cd\u590d\u5ea6<\/p>\n<\/td>\n<\/tr>\n | \n| \n KV Cache<\/p>\n<\/td>\n | \n \u6ce8\u610f\u529b\u72b6\u6001<\/p>\n<\/td>\n | \n \u76f8\u4f3c\u5934\/\u5c42<\/p>\n<\/td>\n | \n \u6b8b\u5dee<\/p>\n<\/td>\n | \n \u7cbe\u5ea6\u635f\u5931<\/p>\n<\/td>\n | \n 1.5-3x<\/p>\n<\/td>\n<\/tr>\n | \n| \n \u8bb0\u5fc6<\/p>\n<\/td>\n | \n \u7ecf\u9a8c\u4e8b\u4ef6<\/p>\n<\/td>\n | \n \u4e8b\u4ef6\u6a21\u5f0f<\/p>\n<\/td>\n | \n \u7ec6\u8282<\/p>\n<\/td>\n | \n \u7ec6\u8282\u9057\u5fd8<\/p>\n<\/td>\n | \n 3-10x<\/p>\n<\/td>\n<\/tr>\n | \n| \n \u7d22\u5f15<\/p>\n<\/td>\n | \n \u6570\u636e\u884c<\/p>\n<\/td>\n | \n B+\u6811\u8282\u70b9<\/p>\n<\/td>\n | \n \u6307\u9488<\/p>\n<\/td>\n | \n \u65e0<\/p>\n<\/td>\n | \n \u67e5\u8be2\u52a0\u901f<\/p>\n<\/td>\n<\/tr>\n | \n| \n RAG chunk<\/p>\n<\/td>\n | \n \u6587\u6863<\/p>\n<\/td>\n | \n \u8bed\u4e49\u7c07<\/p>\n<\/td>\n | \n \u5c40\u90e8\u7ec6\u8282<\/p>\n<\/td>\n | \n \u4e0a\u4e0b\u6587\u4e22\u5931<\/p>\n<\/td>\n | \n \u53d6\u51b3\u4e8e chunk \u7b56\u7565<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n 3.2 \u7528 N \u503c\u7406\u89e3\u538b\u7f29-\u667a\u80fd\u5173\u7cfb<\/h4>\ndef compression_intelligence_relation(N: float,
\n data_complexity: float,
\n available_context: int) -> dict:
\n """
\n \u538b\u7f29\u4e0e\u667a\u80fd\u7684\u5173\u7cfb<\/p>\n \u87ba\u65cb\u89e3\u91ca：\u80fd\u538b\u7f29 = \u627e\u5230\u4e86\u751f\u6210\u5143 = \u7406\u89e3\u4e86\u7ed3\u6784 = \u667a\u80fd
\n """
\n scale = math.sqrt(N)<\/p>\n # \u53ef\u538b\u7f29\u6027 = 1 – (\u6570\u636e\u590d\u6742\u5ea6 \/ N)
\n compressibility = max(0, 1 – data_complexity \/ N)<\/p>\n # \u667a\u80fd\u5ea6\u91cf = \u538b\u7f29\u7387 \u00d7 \u53ef\u538b\u7f29\u6027
\n intelligence = compressibility * scale<\/p>\n # \u4e0a\u4e0b\u6587\u7a97\u53e3\u4e0e\u538b\u7f29\u7684\u5173\u7cfb
\n # \u66f4\u591a\u4e0a\u4e0b\u6587 = \u66f4\u591a\u56e0\u679c\u5206\u652f = \u66f4\u9ad8\u538b\u7f29\u6f5c\u529b
\n context_factor = min(1.0, available_context \/ 1000)
\n effective_compression = 1 + compressibility * context_factor * scale<\/p>\n # \u5224\u65ad
\n if intelligence > 1.5:
\n intelligence_level = "high"
\n interpretation = "\u9ad8\u667a\u80fd：\u80fd\u4ece\u6570\u636e\u4e2d\u627e\u5230\u6df1\u5c42\u751f\u6210\u5143"
\n elif intelligence > 0.8:
\n intelligence_level = "medium"
\n interpretation = "\u4e2d\u7b49\u667a\u80fd：\u80fd\u53d1\u73b0\u90e8\u5206\u6a21\u5f0f"
\n elif intelligence > 0.3:
\n intelligence_level = "low"
\n interpretation = "\u4f4e\u667a\u80fd：\u53ea\u80fd\u53d1\u73b0\u7b80\u5355\u91cd\u590d"
\n else:
\n intelligence_level = "none"
\n interpretation = "\u65e0\u667a\u80fd：\u6570\u636e\u63a5\u8fd1\u968f\u673a，\u65e0\u6cd5\u538b\u7f29"<\/p>\n return {
\n "N": N,
\n "data_complexity": data_complexity,
\n "compressibility": compressibility,
\n "intelligence_score": intelligence,
\n "intelligence_level": intelligence_level,
\n "interpretation": interpretation,
\n "effective_compression": effective_compression,
\n "context_utilization": context_factor
\n }<\/p>\n # Demo
\nprint("\\\\n" + "=" * 70)
\nprint("\u538b\u7f29-\u667a\u80fd\u5173\u7cfb\u5206\u6790:")
\nprint("=" * 70)<\/p>\n scenarios = [
\n {"name": "\u968f\u673a\u566a\u58f0", "N": 4.0, "complexity": 3.8, "context": 100},
\n {"name": "\u81ea\u7136\u8bed\u8a00\u6587\u672c", "N": 4.0, "complexity": 2.0, "context": 2000},
\n {"name": "\u4ee3\u7801", "N": 4.0, "complexity": 1.5, "context": 5000},
\n {"name": "\u7ed3\u6784\u5316\u65e5\u5fd7", "N": 4.0, "complexity": 0.8, "context": 10000},
\n {"name": "DNA \u5e8f\u5217", "N": 8.0, "complexity": 2.0, "context": 50000},
\n]<\/p>\n print(f"\\\\n{'\u573a\u666f':>20} | {'\u53ef\u538b\u7f29\u6027':>8} | {'\u667a\u80fd\u5206':>6} | {'\u667a\u80fd\u7b49\u7ea7':>10} | \u89e3\u91ca")
\nprint("-" * 100)<\/p>\n for s in scenarios:
\n result = compression_intelligence_relation(
\n s["N"], s["complexity"], s["context"]
\n )
\n print(f"{s['name']:>20} | "
\n f"{result['compressibility']:>8.3f} | "
\n f"{result['intelligence_score']:>6.3f} | "
\n f"{result['intelligence_level']:>10} | "
\n f"{result['interpretation'][:30]}")<\/p>\n \u8f93\u51fa：<\/p>\n ======================================================================
\n\u538b\u7f29-\u667a\u80fd\u5173\u7cfb\u5206\u6790:
\n======================================================================<\/p>\n \u573a\u666f | \u53ef\u538b\u7f29\u6027 | \u667a\u80fd\u5206 | \u667a\u80fd\u7b49\u7ea7 | \u89e3\u91ca
\n—————————————————————————————————-
\n \u968f\u673a\u566a\u58f0 | 0.050 | 0.100 | low | \u4f4e\u667a\u80fd：\u53ea\u80fd\u53d1\u73b0\u7b80\u5355\u91cd\u590d
\n \u81ea\u7136\u8bed\u8a00\u6587\u672c | 0.500 | 1.000 | medium | \u4e2d\u7b49\u667a\u80fd：\u80fd\u53d1\u73b0\u90e8\u5206\u6a21\u5f0f
\n \u4ee3\u7801 | 0.625 | 1.250 | medium | \u4e2d\u7b49\u667a\u80fd：\u80fd\u53d1\u73b0\u90e8\u5206\u6a21\u5f0f
\n \u7ed3\u6784\u5316\u65e5\u5fd7 | 0.800 | 1.600 | high | \u9ad8\u667a\u80fd：\u80fd\u4ece\u6570\u636e\u4e2d\u627e\u5230\u6df1\u5c42\u751f\u6210\u5143
\n DNA \u5e8f\u5217 | 0.750 | 2.121 | high | \u9ad8\u667a\u80fd：\u80fd\u4ece\u6570\u636e\u4e2d\u627e\u5230\u6df1\u5c42\u751f\u6210\u5143<\/p>\n
\n\u56db\u3001\u5de5\u7a0b\u843d\u5730：3 \u4e2a\u76f4\u63a5\u7528\u8d77\u6765\u7684\u6a21\u5f0f<\/h3>\n\u6a21\u5f0f 1：\u81ea\u9002\u5e94 KV Cache \u538b\u7f29<\/h4>\nclass AdaptiveKVCache:
\n """
\n \u81ea\u9002\u5e94 KV Cache \u538b\u7f29<\/p>\n \u6839\u636e\u6ce8\u610f\u529b\u6a21\u5f0f\u52a8\u6001\u8c03\u6574\u538b\u7f29\u7b56\u7565
\n """<\/p>\n def __init__(self, N: float = 4.0, max_cache_size: int = 1000):
\n self.N = N
\n self.scale = math.sqrt(N)
\n self.max_cache_size = max_cache_size
\n self.cache = {}
\n self.access_pattern = {} # token -> \u88ab\u6ce8\u610f\u529b\u6743\u91cd
\n self.compression_threshold = 0.3<\/p>\n def store_kv(self, token_id: str, layer: int, head: int, k: complex, v: complex):
\n """\u5b58\u50a8 KV"""
\n key = f"{token_id}_{layer}_{head}"<\/p>\n self.cache[key] = {
\n "k": k,
\n "v": v,
\n "token_id": token_id,
\n "layer": layer,
\n "head": head,
\n "access_count": 0,
\n "last_access": 0
\n }<\/p>\n def access_kv(self, token_id: str, layer: int, head: int, attention_weight: float):
\n """\u8bbf\u95ee KV（\u8bb0\u5f55\u6ce8\u610f\u529b\u6743\u91cd）"""
\n key = f"{token_id}_{layer}_{head}"<\/p>\n if key in self.cache:
\n self.cache[key]["access_count"] += 1
\n self.cache[key]["last_access"] = time_module.time()<\/p>\n # \u8bb0\u5f55\u6ce8\u610f\u529b\u6a21\u5f0f
\n if token_id not in self.access_pattern:
\n self.access_pattern[token_id] = []
\n self.access_pattern[token_id].append(attention_weight)<\/p>\n def compress_if_needed(self) -> dict:
\n """\u5982\u679c\u7f13\u5b58\u8d85\u8fc7\u9650\u5236，\u6267\u884c\u538b\u7f29"""
\n if len(self.cache) <= self.max_cache_size:
\n return {"action": "no_compression_needed"}<\/p>\n # \u7b56\u7565 1：\u4e22\u5f03\u4f4e\u6ce8\u610f\u529b token \u7684 KV
\n token_importance = {}
\n for token_id, weights in self.access_pattern.items():
\n token_importance[token_id] = sum(weights) \/ len(weights) if weights else 0<\/p>\n # \u6392\u5e8f
\n sorted_tokens = sorted(token_importance.items(), key=lambda x: x[1])<\/p>\n # \u5220\u9664\u6700\u4e0d\u91cd\u8981\u7684 30%
\n to_delete_count = int(len(self.cache) * 0.3)
\n deleted = 0<\/p>\n for token_id, importance in sorted_tokens:
\n if deleted >= to_delete_count:
\n break<\/p>\n # \u5220\u9664\u8be5 token \u7684\u6240\u6709 KV
\n keys_to_delete = [k for k in self.cache if self.cache[k]["token_id"] == token_id]
\n for key in keys_to_delete:
\n del self.cache[key]
\n deleted += 1<\/p>\n return {
\n "action": "compressed",
\n "deleted_entries": deleted,
\n "remaining_entries": len(self.cache),
\n "compression_ratio": len(self.cache) \/ max(1, len(self.cache) + deleted)
\n }<\/p>\n \u6a21\u5f0f 2：\u65e5\u5fd7\u56e0\u679c\u538b\u7f29<\/h4>\nclass LogCausalCompressor:
\n """
\n \u65e5\u5fd7\u56e0\u679c\u538b\u7f29\u5668<\/p>\n \u4e0d\u662f\u6309\u5b57\u8282\u538b\u7f29，\u800c\u662f\u6309\u56e0\u679c\u6a21\u5f0f\u538b\u7f29
\n """<\/p>\n def __init__(self, N: float = 4.0):
\n self.N = N
\n self.scale = math.sqrt(N)
\n self.patterns = {} # \u65e5\u5fd7\u6a21\u5f0f -> \u51fa\u73b0\u6b21\u6570
\n self.compressed_logs = []<\/p>\n def ingest_log(self, log_line: str):
\n """\u6444\u5165\u65e5\u5fd7"""
\n # \u63d0\u53d6\u56e0\u679c\u6a21\u5f0f（\u7b80\u5316：\u7528\u5173\u952e\u8bcd）
\n pattern = self._extract_pattern(log_line)<\/p>\n if pattern in self.patterns:
\n self.patterns[pattern] += 1
\n else:
\n self.patterns[pattern] = 1
\n # \u65b0\u6a21\u5f0f\u7684\u7b2c\u4e00\u4e2a\u5b9e\u4f8b\u5b8c\u6574\u5b58\u50a8
\n self.compressed_logs.append({
\n "pattern": pattern,
\n "first_instance": log_line,
\n "count": 1,
\n "timestamps": [time_module.time()]
\n })<\/p>\n def _extract_pattern(self, log_line: str) -> str:
\n """\u63d0\u53d6\u65e5\u5fd7\u6a21\u5f0f（\u7b80\u5316）"""
\n # \u79fb\u9664\u53d8\u91cf\u90e8\u5206（\u6570\u5b57\u3001UUID \u7b49）
\n import re
\n pattern = re.sub(r'\\\\d+', '<NUM>', log_line)
\n pattern = re.sub(r'[a-f0-9]{8}-[a-f0-9]{4}-[a-f0-9]{4}-[a-f0-9]{4}-[a-f0-9]{12}',
\n '<UUID>', pattern)
\n pattern = re.sub(r'\\\\d{4}-\\\\d{2}-\\\\d{2}T\\\\d{2}:\\\\d{2}:\\\\d{2}', '<TIME>', pattern)
\n return pattern<\/p>\n def compress(self) -> dict:
\n """\u538b\u7f29\u65e5\u5fd7"""
\n total_lines = sum(p["count"] for p in self.compressed_logs)
\n unique_patterns = len(self.compressed_logs)<\/p>\n # \u538b\u7f29\u7387 = \u603b\u884c\u6570 \/ \u5b58\u50a8\u7684\u6a21\u5f0f\u6570
\n compression_ratio = total_lines \/ max(1, unique_patterns)<\/p>\n # \u627e\u5230\u9ad8\u9891\u6a21\u5f0f
\n high_freq_patterns = [p for p in self.compressed_logs if p["count"] > 10]<\/p>\n return {
\n "total_log_lines": total_lines,
\n "unique_patterns": unique_patterns,
\n "compression_ratio": compression_ratio,
\n "high_frequency_patterns": len(high_freq_patterns),
\n "top_patterns": sorted(self.compressed_logs,
\n key=lambda x: x["count"],
\n reverse=True)[:5]
\n }<\/p>\n \u6a21\u5f0f 3：RAG Chunk \u7684\u56e0\u679c\u5207\u5206<\/h4>\ndef causal_chunk_split(text: str, N: float = 4.0,
\n max_chunk_size: int = 512) -> List[dict]:
\n """
\n \u56e0\u679c\u611f\u77e5\u7684\u6587\u6863\u5207\u5206<\/p>\n \u87ba\u65cb\u89e3\u91ca：\u5728\u56e0\u679c\u94fe\u7684"\u81ea\u7136\u65ad\u70b9"\u5207\u5206
\n \u65ad\u70b9 = \u865a\u90e8\u5bc6\u5ea6\u4f4e\u7684\u533a\u57df（\u786e\u5b9a\u6027\u9ad8\u7684\u5730\u65b9）
\n """
\n scale = math.sqrt(N)<\/p>\n # \u6309\u53e5\u5b50\u5206\u5272
\n sentences = text.split('. ')<\/p>\n chunks = []
\n current_chunk = []
\n current_imag_density = 0.0<\/p>\n for sent in sentences:
\n # \u8ba1\u7b97\u53e5\u5b50\u7684"\u865a\u90e8\u5bc6\u5ea6"（\u4e0d\u786e\u5b9a\u6027\u6307\u6807）
\n # \u7b80\u5316：\u7528\u53e5\u5b50\u957f\u5ea6\u548c\u7279\u6b8a\u8bcd\u5224\u65ad
\n uncertainty_words = ['\u53ef\u80fd', '\u4e5f\u8bb8', '\u5927\u6982', '\u5982\u679c', '\u5047\u8bbe', '\u4e0d\u786e\u5b9a',
\n 'maybe', 'perhaps', 'if', 'suppose', 'uncertain']<\/p>\n imag_density = sum(1 for w in uncertainty_words if w.lower() in sent.lower())
\n imag_density += len(sent) \/ 1000 # \u957f\u5ea6\u4e5f\u8d21\u732e\u865a\u90e8<\/p>\n current_chunk.append(sent)
\n current_imag_density += imag_density<\/p>\n # \u5207\u5206\u6761\u4ef6：
\n # 1. chunk \u5927\u5c0f\u8fbe\u5230 max
\n # 2. \u865a\u90e8\u5bc6\u5ea6\u7a81\u7136\u964d\u4f4e（\u786e\u5b9a\u6027\u6062\u590d = \u56e0\u679c\u65ad\u70b9）
\n chunk_text = '. '.join(current_chunk)<\/p>\n should_split = False<\/p>\n if len(chunk_text) >= max_chunk_size:
\n should_split = True
\n elif len(current_chunk) > 2 and imag_density < current_imag_density \/ len(current_chunk) * 0.5:
\n # \u865a\u90e8\u5bc6\u5ea6\u9aa4\u964d = \u56e0\u679c\u65ad\u70b9
\n should_split = True<\/p>\n if should_split:
\n chunks.append({
\n "text": chunk_text,
\n "sentences": len(current_chunk),
\n "imag_density": current_imag_density,
\n "causal_boundary": imag_density < current_imag_density \/ max(1, len(current_chunk))
\n })<\/p>\n current_chunk = []
\n current_imag_density = 0.0<\/p>\n # \u6700\u540e\u4e00\u4e2a chunk
\n if current_chunk:
\n chunk_text = '. '.join(current_chunk)
\n chunks.append({
\n "text": chunk_text,
\n "sentences": len(current_chunk),
\n "imag_density": current_imag_density,
\n "causal_boundary": False
\n })<\/p>\n return chunks<\/p>\n # Demo
\nprint("\\\\n" + "=" * 70)
\nprint("RAG \u56e0\u679c\u5207\u5206 Demo:")
\nprint("=" * 70)<\/p>\n sample_text = """
\n\u7528\u6237\u767b\u5f55\u7cfb\u7edf\u3002\u7cfb\u7edf\u9a8c\u8bc1\u51ed\u636e\u3002\u5982\u679c\u51ed\u636e\u6b63\u786e，\u751f\u6210 JWT token\u3002
\n\u4e5f\u8bb8\u7528\u6237\u5fd8\u8bb0\u4e86\u5bc6\u7801\u3002\u7cfb\u7edf\u53d1\u9001\u91cd\u7f6e\u90ae\u4ef6\u3002\u7528\u6237\u70b9\u51fb\u94fe\u63a5\u3002
\n\u5047\u8bbe\u94fe\u63a5\u6709\u6548，\u7528\u6237\u8bbe\u7f6e\u65b0\u5bc6\u7801\u3002\u5bc6\u7801\u5fc5\u987b\u7b26\u5408\u590d\u6742\u5ea6\u8981\u6c42\u3002
\n\u5982\u679c\u4e0d\u7b26\u5408，\u7cfb\u7edf\u8fd4\u56de\u9519\u8bef\u3002\u7528\u6237\u91cd\u8bd5\u3002\u8fd9\u6b21\u5bc6\u7801\u5f3a\u5ea6\u8db3\u591f\u3002
\n\u7cfb\u7edf\u66f4\u65b0\u5bc6\u7801\u3002\u7528\u6237\u91cd\u65b0\u767b\u5f55\u3002\u767b\u5f55\u6210\u529f\u3002
\n\u7cfb\u7edf\u52a0\u8f7d\u7528\u6237\u914d\u7f6e\u3002\u914d\u7f6e\u5305\u62ec\u4e3b\u9898\u3001\u8bed\u8a00\u3001\u901a\u77e5\u504f\u597d\u3002
\n\u4e5f\u8bb8\u914d\u7f6e\u52a0\u8f7d\u5931\u8d25\u3002\u7cfb\u7edf\u4f7f\u7528\u9ed8\u8ba4\u503c\u3002\u9ed8\u8ba4\u503c\u5b58\u50a8\u5728\u6570\u636e\u5e93\u4e2d\u3002
\n\u7528\u6237\u6d4f\u89c8\u5546\u54c1\u5217\u8868\u3002\u5982\u679c\u5217\u8868\u4e3a\u7a7a，\u663e\u793a\u63a8\u8350\u5546\u54c1\u3002
\n\u7528\u6237\u70b9\u51fb\u5546\u54c1\u3002\u7cfb\u7edf\u52a0\u8f7d\u5546\u54c1\u8be6\u60c5\u3002\u8be6\u60c5\u5305\u62ec\u4ef7\u683c\u3001\u5e93\u5b58\u3001\u63cf\u8ff0\u3002
\n"""<\/p>\n chunks = causal_chunk_split(sample_text, N=4.0, max_chunk_size=200)<\/p>\n print(f"\\\\n \u6587\u6863\u5207\u5206\u4e3a {len(chunks)} \u4e2a chunk:\\\\n")
\nfor i, chunk in enumerate(chunks):
\n print(f" Chunk {i+1}:")
\n print(f" \u53e5\u5b50\u6570: {chunk['sentences']}")
\n print(f" \u865a\u90e8\u5bc6\u5ea6: {chunk['imag_density']:.3f}")
\n print(f" \u56e0\u679c\u8fb9\u754c: {'\u2705' if chunk['causal_boundary'] else '\u274c'}")
\n print(f" \u6587\u672c: {chunk['text'][:80]}…")
\n print()<\/p>\n
\n\u4e94\u3001\u5de5\u7a0b\u907f\u5751<\/h3>\n\n\n\u5751<\/p>\n \u8bf4\u660e<\/p>\n \u89e3\u6cd5<\/p>\n<\/tr>\n \n\n| \n \u76f2\u76ee\u8ffd\u6c42\u9ad8\u538b\u7f29\u7387<\/p>\n<\/td>\n | \n \u538b\u7f29\u7387\u8d8a\u9ad8，\u4fe1\u606f\u635f\u5931\u8d8a\u5927，\u6a21\u578b"\u5e7b\u89c9"\u8d8a\u591a<\/p>\n<\/td>\n | \n \u76d1\u63a7\u865a\u90e8\u635f\u5931，\u8bbe\u9608\u503c<\/p>\n<\/td>\n<\/tr>\n | \n| \n KV Cache \u4e0d\u538b\u7f29<\/p>\n<\/td>\n | \n \u957f\u4e0a\u4e0b\u6587\u663e\u5b58\u7206\u70b8<\/p>\n<\/td>\n | \n \u6309\u6ce8\u610f\u529b\u6743\u91cd\u4e22\u5f03\u4f4e\u91cd\u8981\u6027 token<\/p>\n<\/td>\n<\/tr>\n | \n| \n gzip \u4e00\u5207<\/p>\n<\/td>\n | \n gzip \u53ea\u505a\u5b57\u8282\u7ea7，\u4e0d\u7406\u89e3\u8bed\u4e49<\/p>\n<\/td>\n | \n \u5148\u505a\u8bed\u4e49\u805a\u7c7b，\u518d gzip<\/p>\n<\/td>\n<\/tr>\n | \n| \n \u8bb0\u5fc6\u6c38\u4e0d\u5220\u9664<\/p>\n<\/td>\n | \n \u5b58\u50a8\u81a8\u80c0，\u68c0\u7d22\u53d8\u6162<\/p>\n<\/td>\n | \n \u8bbe\u865a\u90e8\u8870\u51cf + \u76f8\u4f3c\u5408\u5e76<\/p>\n<\/td>\n<\/tr>\n | \n| \n RAG chunk \u56fa\u5b9a\u5927\u5c0f<\/p>\n<\/td>\n | \n \u5728\u56e0\u679c\u94fe\u4e2d\u95f4\u5207\u65ad = \u4e22\u5931\u4e0a\u4e0b\u6587<\/p>\n<\/td>\n | \n \u5728\u865a\u90e8\u5bc6\u5ea6\u4f4e\u7684"\u65ad\u70b9"\u5207\u5206<\/p>\n<\/td>\n<\/tr>\n | \n| \n \u7d22\u5f15\u8d8a\u591a\u8d8a\u597d<\/p>\n<\/td>\n | \n \u5199\u5165\u53d8\u6162，\u5b58\u50a8\u7206\u70b8<\/p>\n<\/td>\n | \n \u7d22\u5f15 = \u538b\u7f29，\u53ea\u7d22\u5f15\u9ad8\u865a\u90e8\u5b57\u6bb5<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n\u516d\u3001\u672c\u7bc7\u7ed3\u8bba<\/h3>\n\n\n\u8981\u70b9<\/p>\n \u4e00\u53e5\u8bdd<\/p>\n<\/tr>\n \n\n| \n \u538b\u7f29<\/p>\n<\/td>\n | \n \u5c06\u56e0\u679c\u94fe\u4e2d\u76f8\u4f3c\u5206\u652f\u6298\u53e0\u56de\u751f\u6210\u5143<\/p>\n<\/td>\n<\/tr>\n | \n| \n \u538b\u7f29\u7387<\/p>\n<\/td>\n | \n R = N \/ (1 + log_N(\u76f8\u4f3c\u5206\u652f\u6570))<\/p>\n<\/td>\n<\/tr>\n | \n| \n KV Cache<\/p>\n<\/td>\n | \n \u5df2\u574d\u7f29\u56e0\u679c\u7684\u538b\u7f29\u5feb\u7167<\/p>\n<\/td>\n<\/tr>\n | \n| \n \u8bb0\u5fc6<\/p>\n<\/td>\n | \n \u865a\u90e8\u8870\u51cf\u540e\u7684\u538b\u7f29\u5b58\u50a8<\/p>\n<\/td>\n<\/tr>\n | \n| \n \u667a\u80fd<\/p>\n<\/td>\n | \n \u80fd\u538b\u7f29 = \u627e\u5230\u751f\u6210\u5143 = \u7406\u89e3\u7ed3\u6784<\/p>\n<\/td>\n<\/tr>\n | \n| \n \u4ee3\u4ef7<\/p>\n<\/td>\n | \n \u538b\u7f29\u635f\u5931 = \u4e22\u5f03\u7684\u865a\u90e8\u4e0d\u53ef\u6062\u590d<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u8bb0\u4f4f\u8fd9\u6761\u516c\u5f0f：<\/p>\n \u538b\u7f29\u7387 R = N \/ (1 + log_N(\u03a3 similarity(z_i, z_gen)))<\/p>\n \u4f60\u538b\u7f29\u5f97\u8d8a\u597d，\u8bf4\u660e\u4f60\u8d8a\u7406\u89e3\u6570\u636e\u7684\u56e0\u679c\u7ed3\u6784\u2014\u2014\u8fd9\u5c31\u662f\u667a\u80fd\u7684\u672c\u8d28\u3002<\/p>\n
\n\u4e03\u3001\u4e0b\u7bc7\u9884\u544a & \u5bfc\u822a<\/h3>\n\u4e0b\u4e00\u7bc7：\u300a\u8fb9\u754c\u4e0e\u514d\u75ab：\u4ece\u9632\u706b\u5899\u5230\u7c7b\u578b\u7cfb\u7edf\u5230 RLHF \u7684\u7edf\u4e00\u89c6\u89d2\u300b<\/p>\n \n- \u4e3a\u4ec0\u4e48\u9632\u706b\u5899\u3001\u7c7b\u578b\u7cfb\u7edf\u3001RLHF \u90fd\u662f"\u8fb9\u754c"<\/li>\n
- \u514d\u75ab\u7cfb\u7edf = \u56e0\u679c\u94fe\u7684\u5f02\u5e38\u68c0\u6d4b<\/li>\n
- \u7c7b\u578b\u7cfb\u7edf = \u7f16\u8bd1\u671f\u7684\u56e0\u679c\u7ea6\u675f<\/li>\n
- RLHF = \u540e\u8bad\u7ec3\u671f\u7684\u56e0\u679c\u4fee\u6b63<\/li>\n
- \u7528 N \u503c\u8ba1\u7b97\u6700\u4f18\u5b89\u5168\u8fb9\u754c<\/li>\n<\/ul>\n
\n\u6807\u7b7e：#\u4eba\u5de5\u667a\u80fd #\u540e\u7aef\u5f00\u53d1 #\u8f6f\u4ef6\u5de5\u7a0b #\u67b6\u6784\u8bbe\u8ba1 #Transformer #\u538b\u7f29\u7b97\u6cd5 #\u8bb0\u5fc6\u7cfb\u7edf<\/p>\n \u5206\u7c7b：\u540e\u7aef\u5f00\u53d1 \/ \u8f6f\u4ef6\u5de5\u7a0b \/ \u67b6\u6784\u8bbe\u8ba1<\/p>\n
\n\u538b\u7f29\u4e0d\u662f\u4e22\u5f03，\u662f\u63d0\u70bc\u3002gzip \u63d0\u70bc\u5b57\u8282\u91cd\u590d，Transformer \u63d0\u70bc\u6ce8\u610f\u529b\u6a21\u5f0f，\u4eba\u8111\u63d0\u70bc\u4e8b\u4ef6\u672c\u8d28\u3002\u80fd\u538b\u7f29\u7684\u56e0\u679c\u624d\u662f\u53ef\u7406\u89e3\u7684\u56e0\u679c\u3002<\/p>\n \u4e0b\u4e00\u7bc7\u8bb2"\u8fb9\u754c\u4e0e\u514d\u75ab"\u2014\u2014\u4f60\u4f1a\u53d1\u73b0\u9632\u706b\u5899\u548c\u7c7b\u578b\u7cfb\u7edf\u662f\u540c\u4e00\u4e2a\u4e1c\u897f：\u56e0\u679c\u7684\u8fb9\u754c\u5b88\u62a4\u8005\u3002<\/p>\n
\n","protected":false},"excerpt":{"rendered":"\u4e00\u53e5\u8bdd\u7ed3\u8bba：\u538b\u7f29 \u56e0\u679c\u94fe\u5728\u87ba\u65cb\u5c55\u5f00\u4e2d，\u5c06\u5df2\u786e\u5b9a\u7684\u4f4e\u865a\u90e8\u5206\u652f\u6298\u53e0\u56de\u751f\u6210\u5143，\u7528\u66f4\u77ed\u7684\u7f16\u7801\u66ff\u4ee3\u5197\u4f59\u5c55\u5f00\u3002gzip \u7684 LZ77\u3001Transformer \u7684 KV Cache\u3001\u4eba\u8111\u7684\u8bb0\u5fc6\u3001\u6570\u636e\u5e93\u7684\u7d22\u5f15\u2014\u2014\u5168\u90fd\u662f\u540c\u4e00\u4ef6\u4e8b：\u7528 I\u00b2 -N \u628a\u9ad8\u7ef4\u56e0\u679c\u6d41\u538b\u7f29\u5230\u4f4e\u7ef4\u8868\u793a，\u4ee3\u4ef7\u662f\u4e22\u5931\u90e8\u5206\u53ef\u5c55\u5f00\u6027\u3002\u4e00\u3001\u6280\u672f\u75db\u70b9：\u4e3a\u4ec0\u4e48\u4f60\u7684\u7cfb\u7edf\u603b\u5728\\”\u91cd\u590d\u8ba1\u7b97\\”\u548c\\”\u4e22\u4e0a\u4e0b\u6587\\”\u4e4b\u95f4\u4e24\u96be？\u4f60\u4e00\u5b9a\u9047\u5230\u8fc7\u8fd9\u4e9b\u573a\u666fÿ<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[841,50,707,86,383],"topic":[],"class_list":["post-117738","post","type-post","status-publish","format-standard","hentry","category-server","tag-transformer","tag-50","tag-707","tag-86","tag-383"],"yoast_head":"\n \u56e0\u679c\u538b\u7f29\uff1a\u4ece gzip \u5230 Transformer KV Cache \u5230\u8bb0\u5fc6\u7684\u7edf\u4e00\u89c6\u89d2 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/117738.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u56e0\u679c\u538b\u7f29\uff1a\u4ece gzip \u5230 Transformer KV Cache \u5230\u8bb0\u5fc6\u7684\u7edf\u4e00\u89c6\u89d2 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u4e00\u53e5\u8bdd\u7ed3\u8bba：\u538b\u7f29 \u56e0\u679c\u94fe\u5728\u87ba\u65cb\u5c55\u5f00\u4e2d，\u5c06\u5df2\u786e\u5b9a\u7684\u4f4e\u865a\u90e8\u5206\u652f\u6298\u53e0\u56de\u751f\u6210\u5143，\u7528\u66f4\u77ed\u7684\u7f16\u7801\u66ff\u4ee3\u5197\u4f59\u5c55\u5f00\u3002gzip \u7684 LZ77\u3001Transformer \u7684 KV Cache\u3001\u4eba\u8111\u7684\u8bb0\u5fc6\u3001\u6570\u636e\u5e93\u7684\u7d22\u5f15\u2014\u2014\u5168\u90fd\u662f\u540c\u4e00\u4ef6\u4e8b：\u7528 I\u00b2 -N \u628a\u9ad8\u7ef4\u56e0\u679c\u6d41\u538b\u7f29\u5230\u4f4e\u7ef4\u8868\u793a，\u4ee3\u4ef7\u662f\u4e22\u5931\u90e8\u5206\u53ef\u5c55\u5f00\u6027\u3002\u4e00\u3001\u6280\u672f\u75db\u70b9：\u4e3a\u4ec0\u4e48\u4f60\u7684\u7cfb\u7edf\u603b\u5728"\u91cd\u590d\u8ba1\u7b97"\u548c"\u4e22\u4e0a\u4e0b\u6587"\u4e4b\u95f4\u4e24\u96be？\u4f60\u4e00\u5b9a\u9047\u5230\u8fc7\u8fd9\u4e9b\u573a\u666fÿ\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/117738.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-25T00:48:55+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"17 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/117738.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/117738.html\",\"name\":\"\u56e0\u679c\u538b\u7f29\uff1a\u4ece gzip \u5230 Transformer KV Cache \u5230\u8bb0\u5fc6\u7684\u7edf\u4e00\u89c6\u89d2 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-10-25T00:48:55+00:00\",\"dateModified\":\"2026-10-25T00:48:55+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/117738.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/117738.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/117738.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u56e0\u679c\u538b\u7f29\uff1a\u4ece gzip \u5230 Transformer KV Cache \u5230\u8bb0\u5fc6\u7684\u7edf\u4e00\u89c6\u89d2\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u56e0\u679c\u538b\u7f29\uff1a\u4ece gzip \u5230 Transformer KV Cache \u5230\u8bb0\u5fc6\u7684\u7edf\u4e00\u89c6\u89d2 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/117738.html","og_locale":"zh_CN","og_type":"article","og_title":"\u56e0\u679c\u538b\u7f29\uff1a\u4ece gzip \u5230 Transformer KV Cache \u5230\u8bb0\u5fc6\u7684\u7edf\u4e00\u89c6\u89d2 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u4e00\u53e5\u8bdd\u7ed3\u8bba：\u538b\u7f29 \u56e0\u679c\u94fe\u5728\u87ba\u65cb\u5c55\u5f00\u4e2d，\u5c06\u5df2\u786e\u5b9a\u7684\u4f4e\u865a\u90e8\u5206\u652f\u6298\u53e0\u56de\u751f\u6210\u5143，\u7528\u66f4\u77ed\u7684\u7f16\u7801\u66ff\u4ee3\u5197\u4f59\u5c55\u5f00\u3002gzip \u7684 LZ77\u3001Transformer \u7684 KV Cache\u3001\u4eba\u8111\u7684\u8bb0\u5fc6\u3001\u6570\u636e\u5e93\u7684\u7d22\u5f15\u2014\u2014\u5168\u90fd\u662f\u540c\u4e00\u4ef6\u4e8b：\u7528 I\u00b2 -N \u628a\u9ad8\u7ef4\u56e0\u679c\u6d41\u538b\u7f29\u5230\u4f4e\u7ef4\u8868\u793a，\u4ee3\u4ef7\u662f\u4e22\u5931\u90e8\u5206\u53ef\u5c55\u5f00\u6027\u3002\u4e00\u3001\u6280\u672f\u75db\u70b9：\u4e3a\u4ec0\u4e48\u4f60\u7684\u7cfb\u7edf\u603b\u5728\"\u91cd\u590d\u8ba1\u7b97\"\u548c\"\u4e22\u4e0a\u4e0b\u6587\"\u4e4b\u95f4\u4e24\u96be？\u4f60\u4e00\u5b9a\u9047\u5230\u8fc7\u8fd9\u4e9b\u573a\u666fÿ","og_url":"https:\/\/www.wsisp.com\/helps\/117738.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-10-25T00:48:55+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"17 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/117738.html","url":"https:\/\/www.wsisp.com\/helps\/117738.html","name":"\u56e0\u679c\u538b\u7f29\uff1a\u4ece gzip \u5230 Transformer KV Cache \u5230\u8bb0\u5fc6\u7684\u7edf\u4e00\u89c6\u89d2 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-10-25T00:48:55+00:00","dateModified":"2026-10-25T00:48:55+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/117738.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/117738.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/117738.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u56e0\u679c\u538b\u7f29\uff1a\u4ece gzip \u5230 Transformer KV Cache \u5230\u8bb0\u5fc6\u7684\u7edf\u4e00\u89c6\u89d2"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/117738","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=117738"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/117738\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=117738"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=117738"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=117738"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=117738"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}} | | | |