Spaces:
Paused
Paused
File size: 10,452 Bytes
751ad26 | 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 | from __future__ import annotations
import numpy as np
from .decode_reference import decode_page
from .modes.m1_lut import dequantize_group_lut
from .modes.m2_key_sketch import segment_ids_for_token_count
from .modes.m4_key_project import fixed_project_basis
from .modes.m3_escape import decode_escape_payload
from .modes.turbo3 import fwht_last_dim
from .page_format import load_group_words
from .packing import unpack_bits
from .types import EncodedPage
def _pad_query(query_slice: np.ndarray, padded_head_dim: int) -> np.ndarray:
query = np.asarray(query_slice, dtype=np.float32)
if query.ndim != 1:
raise ValueError("query_slice must have shape [head_dim]")
if query.shape[0] > padded_head_dim:
raise ValueError("query head_dim exceeds padded_head_dim")
if query.shape[0] == padded_head_dim:
return query
return np.pad(query, (0, padded_head_dim - query.shape[0]), mode="constant")
def softmax(logits: np.ndarray) -> np.ndarray:
values = np.asarray(logits, dtype=np.float32)
shifted = values - np.max(values)
weights = np.exp(shifted)
return weights / np.sum(weights)
def score_page_ref(query_slice: np.ndarray, page: EncodedPage) -> np.ndarray:
header = page.header
query = _pad_query(query_slice, header.padded_head_dim)
if header.mode_default == "M3":
if page.escape_payload is None:
raise ValueError("escape payload is missing")
dense = decode_escape_payload(page.escape_payload, head_dim=header.head_dim, scales=page.escape_scales)
return dense @ query_slice.astype(np.float32)
if header.mode_default == "M2":
if page.m2_sketch is None or page.m2_basis is None:
raise ValueError("M2 page is missing sketch payload")
query_groups = query.reshape(header.num_groups, header.group_size)
logits = np.zeros(header.token_count, dtype=np.float32)
for group_index in range(header.num_groups):
group_mean = None if page.m2_mean is None else page.m2_mean[group_index].astype(np.float32)
group_basis = page.m2_basis[group_index].astype(np.float32)
if group_basis.ndim == 2:
q_proj = group_basis @ query_groups[group_index]
logits += page.m2_sketch[:, group_index, :].astype(np.float32) @ q_proj.astype(np.float32)
if group_mean is not None:
logits += np.dot(group_mean, query_groups[group_index]).astype(np.float32)
continue
segment_ids = segment_ids_for_token_count(header.token_count, int(group_basis.shape[0]))
q_proj = np.einsum("srg,g->sr", group_basis, query_groups[group_index])
logits += np.einsum("tr,tr->t", page.m2_sketch[:, group_index, :].astype(np.float32), q_proj[segment_ids])
if group_mean is not None:
logits += group_mean[segment_ids].astype(np.float32) @ query_groups[group_index]
return logits
if header.mode_default == "M4":
if page.m2_sketch is None or page.m2_mean is None:
raise ValueError("M4 page is missing projected payload")
query_groups = query.reshape(header.num_groups, header.group_size)
logits = np.zeros(header.token_count, dtype=np.float32)
for group_index in range(header.num_groups):
basis = (
np.asarray(page.m2_basis[group_index], dtype=np.float32)
if page.m2_basis is not None
else fixed_project_basis(header.group_size, int(page.m2_sketch.shape[-1]), header.project_basis)
)
q_proj = basis @ query_groups[group_index]
logits += page.m2_sketch[:, group_index, :].astype(np.float32) @ q_proj.astype(np.float32)
logits += np.dot(page.m2_mean[group_index].astype(np.float32), query_groups[group_index]).astype(np.float32)
return logits
if header.mode_default == "T3":
if page.payload is None or page.scales is None or page.codebooks is None:
raise ValueError("T3 page is missing payload or correction metadata")
rotated_query_groups = fwht_last_dim(query.reshape(header.num_groups, header.group_size))
logits = np.zeros(header.token_count, dtype=np.float32)
centroids = np.asarray(page.codebooks, dtype=np.float32)
for group_index in range(header.num_groups):
words = load_group_words(page, group_index)
codes_u8 = unpack_bits(words, header.bits, header.group_size).astype(np.int64, copy=False)
corrected = centroids[codes_u8] * page.scales[:, group_index].astype(np.float32)[:, None]
logits += corrected @ rotated_query_groups[group_index]
return logits
if page.payload is None:
raise ValueError(f"{header.mode_default} page is missing payload")
query_groups = query.reshape(header.num_groups, header.group_size)
query_group_sums = query_groups.sum(axis=-1)
logits = np.zeros(header.token_count, dtype=np.float32)
for group_index in range(header.num_groups):
words = load_group_words(page, group_index)
codes_u8 = unpack_bits(words, header.bits, header.group_size)
qg = query_groups[group_index]
if header.mode_default == "M1":
if page.codebooks is None:
raise ValueError("M1 page is missing codebooks")
group = dequantize_group_lut(codes_u8, codebook=np.asarray(page.codebooks[group_index], dtype=np.float32))
logits += group @ qg
continue
if page.scales is None:
raise ValueError("M0 page is missing scales")
codes = codes_u8.astype(np.float32)
scales = page.scales[:, group_index].astype(np.float32)
if header.quant_scheme == "affine":
if page.bias is None:
raise ValueError("affine pages require bias metadata")
int_dot = codes @ qg
bias = page.bias[:, group_index].astype(np.float32)
logits += scales * int_dot + bias * query_group_sums[group_index]
continue
zero_point = (1 << (header.bits - 1)) - 1
logits += scales * ((codes - zero_point) @ qg)
return logits
def mix_page_ref(attn_weights: np.ndarray, page: EncodedPage, out_acc: np.ndarray | None = None) -> np.ndarray:
header = page.header
weights = np.asarray(attn_weights, dtype=np.float32)
if weights.shape != (header.token_count,):
raise ValueError("attn_weights must have shape [token_count]")
output = np.zeros(header.padded_head_dim, dtype=np.float32) if out_acc is None else np.asarray(out_acc, dtype=np.float32)
if output.shape != (header.padded_head_dim,):
raise ValueError("out_acc must have shape [padded_head_dim]")
if header.mode_default == "M3":
if page.escape_payload is None:
raise ValueError("escape payload is missing")
dense = decode_escape_payload(page.escape_payload, head_dim=header.head_dim, scales=page.escape_scales)
output[: header.head_dim] += weights @ dense
return output[: header.head_dim].copy()
if header.mode_default in {"M2", "M4"}:
raise ValueError(f"{header.mode_default} is only supported for key scoring in this phase")
if header.mode_default == "T3":
if page.payload is None or page.scales is None or page.codebooks is None:
raise ValueError("T3 page is missing payload or correction metadata")
centroids = np.asarray(page.codebooks, dtype=np.float32)
for group_index in range(header.num_groups):
words = load_group_words(page, group_index)
codes_u8 = unpack_bits(words, header.bits, header.group_size).astype(np.int64, copy=False)
rotated_group = centroids[codes_u8] * page.scales[:, group_index].astype(np.float32)[:, None]
group = fwht_last_dim(rotated_group)
start = group_index * header.group_size
end = start + header.group_size
output[start:end] += weights @ group
return output[: header.head_dim].copy()
if page.payload is None:
raise ValueError(f"{header.mode_default} page is missing payload")
for group_index in range(header.num_groups):
words = load_group_words(page, group_index)
codes_u8 = unpack_bits(words, header.bits, header.group_size)
if header.mode_default == "M1":
if page.codebooks is None:
raise ValueError("M1 page is missing codebooks")
group = dequantize_group_lut(codes_u8, codebook=np.asarray(page.codebooks[group_index], dtype=np.float32))
else:
if page.scales is None:
raise ValueError("M0 page is missing scales")
codes = codes_u8.astype(np.float32)
scales = page.scales[:, group_index].astype(np.float32)[:, None]
if header.quant_scheme == "affine":
if page.bias is None:
raise ValueError("affine pages require bias metadata")
group = scales * codes + page.bias[:, group_index].astype(np.float32)[:, None]
else:
zero_point = (1 << (header.bits - 1)) - 1
group = scales * (codes - zero_point)
start = group_index * header.group_size
end = start + header.group_size
output[start:end] += weights @ group
return output[: header.head_dim].copy()
def explicit_dequantized_score(query_slice: np.ndarray, page: EncodedPage) -> np.ndarray:
dense = decode_page(page)
query = np.asarray(query_slice, dtype=np.float32)
return dense @ query
def explicit_dequantized_mix(attn_weights: np.ndarray, page: EncodedPage) -> np.ndarray:
dense = decode_page(page)
weights = np.asarray(attn_weights, dtype=np.float32)
return weights @ dense
def run_attention_reference(query_slice: np.ndarray, key_page: EncodedPage, value_page: EncodedPage) -> tuple[np.ndarray, np.ndarray]:
logits = score_page_ref(query_slice, key_page)
weights = softmax(logits)
output = mix_page_ref(weights, value_page)
return logits, output
def explicit_dequantized_attention(query_slice: np.ndarray, key_page: EncodedPage, value_page: EncodedPage) -> tuple[np.ndarray, np.ndarray]:
logits = explicit_dequantized_score(query_slice, key_page)
weights = softmax(logits)
output = explicit_dequantized_mix(weights, value_page)
return logits, output
|