Ayaka157 commited on
Commit
cfad348
·
verified ·
1 Parent(s): c21fd57

Upload retention_fast.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. retention_fast.py +180 -0
retention_fast.py ADDED
@@ -0,0 +1,180 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Batched-chunk chunkwise retention — mathematically identical to the reference
2
+ per-chunk loop in ``retnet_hf/modeling_retnet.py``, but with ~30x fewer kernel
3
+ launches.
4
+
5
+ Reference (per-chunk loop, 4 matmuls per chunk):
6
+
7
+ for i in chunks:
8
+ inner = (q_i @ k_i.T * inner_mask) @ v_i
9
+ cross = (q_i * cross_decay) @ state
10
+ out_i = inner + cross
11
+ state = chunk_decay * state + k_i.T @ (v_i * value_decay)
12
+
13
+ Batched form (3 matmuls total, independent of the chunk count):
14
+
15
+ 1. U_i = k_i.T @ (v_i * value_decay) one batched matmul
16
+ 2. S_prev_i = scan(U, chunk_decay, state) tiny sequential loop on
17
+ [B,H,kd,vd] tensors only
18
+ 3. inner = (q_r @ k_r.T * inner_mask) @ v_r one batched matmul
19
+ cross = (q_r * cross_decay) @ S_prev one batched matmul
20
+
21
+ The scan in step 2 is deliberately *not* replaced by a cumsum: with
22
+ ``chunk_decay = gamma^C`` and C=256, ``gamma^-j`` overflows fp32 for the fast
23
+ heads, while the log-depth pairwise scan costs more than it saves at 8 chunks.
24
+ """
25
+ from __future__ import annotations
26
+
27
+ import torch
28
+
29
+
30
+ _INDEX_CACHE: dict = {}
31
+
32
+
33
+ def _index_tables(chunk_size: int, device):
34
+ """``causal`` / ``expo`` / ``rel`` 只依赖 chunk_size,缓存掉。
35
+
36
+ 原实现每次调用都重算 ``arange``/``sub``/``ge``/``clamp``,是 4 次无用的小
37
+ 内核发射;而本算子的成本**几乎全是发射延迟**(见下)。
38
+ """
39
+ key = (int(chunk_size), str(device))
40
+ tables = _INDEX_CACHE.get(key)
41
+ if tables is None:
42
+ rel = torch.arange(chunk_size, device=device)
43
+ delta = rel[:, None] - rel[None, :]
44
+ causal = delta >= 0
45
+ expo = delta.clamp_min(0).to(torch.float32)
46
+ tables = (causal, expo, rel.to(torch.float32))
47
+ _INDEX_CACHE[key] = tables
48
+ return tables
49
+
50
+
51
+ def _decay_tables(gamma: torch.Tensor, chunk_size: int, dtype: torch.dtype,
52
+ device: torch.device):
53
+ """gamma: [H] (fp32)。返回 inner_mask [H,C,C], cross_decay [H,C],
54
+ value_decay [H,C], chunk_decay [H],均为 ``dtype``。
55
+
56
+ 块内掩码用**对数空间** ``exp(log γ · delta)``,这是唯一在本模型门控范围
57
+ 两端都安全的写法:
58
+
59
+ * 因式分解的 ``γ^i · γ^-j`` 一旦需要 ``γ^-255`` 就溢出 fp32。
60
+ ``retnet_272m_rp`` 的 GLA 门控最小到 0.668,``0.668**-255 = 1e44 > 3.4e38``
61
+ → Inf → 前向 NaN(``torch.where`` 只丢结果,不阻止 Inf 被算出来)。
62
+ * 朴素 ``γ ** delta`` 对 γ<1 安全(下溢到 0,那里本来就该是 0),
63
+ 但对 γ>1 的门控不成立;对数形式两端都对。
64
+
65
+ NOTE(2026-10-10): 这个算子只有 26 万个元素却要花 **410 µs** —— 成本几乎全是
66
+ 「22 个相互依赖的小内核」的发射延迟(Windows 上约 18 µs/次)。所以这里的
67
+ 优化目标一律是**减少发射次数**,而不是减少计算量:
68
+ * 索引张量进缓存(省 4 次发射);
69
+ * 三张一维表复用同一个 ``γ^j``:[H,C]:``cross = γ^j·γ``、
70
+ ``value = flip(γ^j)``、``chunk = γ^j[-1]·γ``,
71
+ 把原来的 3 次 ``exp`` 压成 1 次。
72
+ """
73
+ causal, expo, relf = _index_tables(chunk_size, device)
74
+ g = gamma.to(device=device, dtype=torch.float32)
75
+ # γ 可以合法地取到 0(门控被训练成「完全遗忘」),此时 log(0) = -inf,
76
+ # 而对角线 delta=0 会让 exp(-inf * 0) = exp(NaN) —— torch.where 只丢结果、
77
+ # 不阻止 NaN 被算出来,于是整层 NaN。夹一个正下限后
78
+ # exp(logγ·0) = 1、exp(logγ·k) 下溢到 0,正好是参考实现 gamma**delta 的
79
+ # 语义(0**0 = 1、0**k = 0),两条路径在 γ=0 处重新等价。
80
+ log_gamma = torch.log(g.clamp_min(torch.finfo(torch.float32).tiny))
81
+
82
+ inner_mask = torch.exp(log_gamma[:, None, None] * expo[None])
83
+ inner_mask = torch.where(causal, inner_mask,
84
+ torch.zeros((), device=device, dtype=torch.float32))
85
+
86
+ pow1d = torch.exp(log_gamma[:, None] * relf[None, :]) # [H,C] = γ^j
87
+ cross_decay = pow1d * g[:, None] # γ^(j+1)
88
+ value_decay = pow1d.flip(-1) # γ^(C-1-j)
89
+ chunk_decay = pow1d[:, -1] * g # γ^C
90
+
91
+ return (inner_mask.to(dtype), cross_decay.to(dtype),
92
+ value_decay.to(dtype), chunk_decay.to(dtype))
93
+
94
+
95
+ def chunkwise_batched(q, k, v, gamma, state, chunk_size):
96
+ """Drop-in replacement for ``MultiScaleRetention._chunkwise_retention``.
97
+
98
+ q, k: [B,H,L,kd] v: [B,H,L,vd] gamma: [H] fp32
99
+ state: [B,H,kd,vd] or None -> (output [B,H,L,vd], new_state)
100
+ """
101
+ B, H, L, kd = q.shape
102
+ vd = v.shape[-1]
103
+ if L % chunk_size != 0:
104
+ raise ValueError(f"seq_len {L} not a multiple of chunk_size {chunk_size}")
105
+ nc = L // chunk_size
106
+ C = chunk_size
107
+ dev = q.device
108
+
109
+ inner_mask, cross_decay, value_decay, chunk_decay = _decay_tables(
110
+ gamma.to(device=dev, dtype=torch.float32), C, q.dtype, dev)
111
+
112
+ has_state = state is not None
113
+ if state is None:
114
+ state = torch.zeros(B, H, kd, vd, device=dev, dtype=q.dtype)
115
+
116
+ # [B,H,nc,C,*] views — no copy (L is contiguous)
117
+ qr = q.view(B, H, nc, C, kd)
118
+ kr = k.view(B, H, nc, C, kd)
119
+ vr = v.view(B, H, nc, C, vd)
120
+
121
+ # (1) per-chunk state contributions, one batched matmul over all chunks
122
+ U = torch.matmul(kr.transpose(-1, -2),
123
+ vr * value_decay[None, :, None, :, None]) # [B,H,nc,kd,vd]
124
+
125
+ # (2) exclusive scan -> S_prev[i] = state before chunk i
126
+ #
127
+ # NOTE(2026-10-10): 原来的写法是 nc 次串行 ``s = decay*s + U[:,:,i]``
128
+ # 外加一次 ``torch.stack`` —— nc=16 时 33 次内核发射、实测 **591 µs**,
129
+ # 而每个参与的张量只有 0.5 MB:几乎全部是发射延迟。
130
+ #
131
+ # 关键观察是 ``decay = chunk_decay`` 在整个循环里**是常量**(它是 [H] 的
132
+ # 每头常数,不随 chunk 变),于是有闭式解
133
+ #
134
+ # S_prev[i] = d^i · state + Σ_{j<i} d^(i-1-j) · U_j
135
+ #
136
+ # 这正是对 ``T[h,i,j] = d_h^(i-1-j)``(下三角)的一次批量矩阵乘:
137
+ #
138
+ # S_all = T[None] @ U.reshape(B,H,nc,kd*vd) [B,H,nc+1,kd*vd]
139
+ #
140
+ # 多出来的一行 i=nc 恰好就是 new_state。**所有指数非负**,因此不存在
141
+ # §2.1 那种 ``γ^-j`` 溢出——d 再小也只是下溢到 0,那本就代表历史被忘光。
142
+ #
143
+ # NOTE(2026-10-10): 之前试过对数深度(Hillis-Steele)扫描把启动数从 2·nc
144
+ # 降到 ~2·log2(nc),**反而慢 2.8 倍**(3004 vs 1078 ms),内核数还从
145
+ # 10116 涨到 11472 —— 因为每轮都要 ``torch.cat`` 物化一个
146
+ # [B,H,nc,kd,vd] 张量,既贵又破坏 inductor 的融合。这里的闭式解不物化任何
147
+ # 额外的大张量(``T`` 只有 H·(nc+1)·nc 个元素),所以是另一回事。
148
+ # Collected in a list and stacked: writing into a torch.empty() buffer
149
+ # with slice assignment makes autograd emit CopySlices and leaves the
150
+ # untouched-in-backward buffer as uninitialised memory.
151
+ # 同样要夹下限:expo_t 在 (i=0, j=0) 处为 -1,clamp_min(0) 后是 0,
152
+ # 若 log_d 是 -inf 则 -inf*0 = NaN 会留在 T 的对角线上。
153
+ log_d = (torch.log(gamma.to(device=dev, dtype=torch.float32)
154
+ .clamp_min(torch.finfo(torch.float32).tiny))
155
+ * float(chunk_size))
156
+ idx = torch.arange(nc + 1, device=dev)
157
+ expo_t = idx[:, None] - torch.arange(nc, device=dev)[None, :] - 1
158
+ valid_t = expo_t >= 0
159
+ T = torch.exp(log_d[:, None, None] * expo_t.clamp_min(0).to(torch.float32)[None])
160
+ T = torch.where(valid_t[None], T, torch.zeros((), device=dev, dtype=torch.float32))
161
+ T = T.to(q.dtype) # [H,nc+1,nc]
162
+
163
+ U_flat = U.reshape(B, H, nc, kd * vd)
164
+ S_all = torch.matmul(T[None], U_flat) # [B,H,nc+1,kd*vd]
165
+ if has_state:
166
+ t_state = torch.exp(log_d[:, None] * idx.to(torch.float32)[None]).to(q.dtype)
167
+ S_all = S_all + t_state[None, :, :, None] * state.reshape(B, H, 1, kd * vd)
168
+
169
+ S_prev = S_all[:, :, :nc].reshape(B, H, nc, kd, vd)
170
+ new_state = S_all[:, :, nc].reshape(B, H, kd, vd)
171
+
172
+ # (3) inner chunk, one batched matmul
173
+ inner = torch.matmul(
174
+ torch.matmul(qr, kr.transpose(-1, -2)) * inner_mask[None, :, None, :, :], vr)
175
+
176
+ # (4) cross chunk, one batched matmul
177
+ cross = torch.matmul(qr * cross_decay[None, :, None, :, None], S_prev)
178
+
179
+ out = (inner + cross).reshape(B, H, L, vd)
180
+ return out, new_state