RetNet-0.5B-v21 / retention_fast.py
Ayaka157's picture
Upload folder using huggingface_hub
a9e39bc verified
Raw History Blame Contribute Delete
8.7 kB
"""Batched-chunk chunkwise retention — mathematically identical to the reference
per-chunk loop in ``retnet_hf/modeling_retnet.py``, but with ~30x fewer kernel
launches.
Reference (per-chunk loop, 4 matmuls per chunk):
for i in chunks:
inner = (q_i @ k_i.T * inner_mask) @ v_i
cross = (q_i * cross_decay) @ state
out_i = inner + cross
state = chunk_decay * state + k_i.T @ (v_i * value_decay)
Batched form (3 matmuls total, independent of the chunk count):
1. U_i = k_i.T @ (v_i * value_decay) one batched matmul
2. S_prev_i = scan(U, chunk_decay, state) tiny sequential loop on
[B,H,kd,vd] tensors only
3. inner = (q_r @ k_r.T * inner_mask) @ v_r one batched matmul
cross = (q_r * cross_decay) @ S_prev one batched matmul
The scan in step 2 is deliberately *not* replaced by a cumsum: with
``chunk_decay = gamma^C`` and C=256, ``gamma^-j`` overflows fp32 for the fast
heads, while the log-depth pairwise scan costs more than it saves at 8 chunks.
"""
from __future__ import annotations
import torch
_INDEX_CACHE: dict = {}
def _index_tables(chunk_size: int, device):
"""``causal`` / ``expo`` / ``rel`` 只依赖 chunk_size,缓存掉。
原实现每次调用都重算 ``arange``/``sub``/``ge``/``clamp``,是 4 次无用的小
内核发射;而本算子的成本**几乎全是发射延迟**(见下)。
"""
key = (int(chunk_size), str(device))
tables = _INDEX_CACHE.get(key)
if tables is None:
rel = torch.arange(chunk_size, device=device)
delta = rel[:, None] - rel[None, :]
causal = delta >= 0
expo = delta.clamp_min(0).to(torch.float32)
tables = (causal, expo, rel.to(torch.float32))
_INDEX_CACHE[key] = tables
return tables
def _decay_tables(gamma: torch.Tensor, chunk_size: int, dtype: torch.dtype,
device: torch.device):
"""gamma: [H] (fp32)。返回 inner_mask [H,C,C], cross_decay [H,C],
value_decay [H,C], chunk_decay [H],均为 ``dtype``。
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
两端都安全的写法:
* 因式分解的 ``γ^i · γ^-j`` 一旦需要 ``γ^-255`` 就溢出 fp32。
``retnet_272m_rp`` 的 GLA 门控最小到 0.668,``0.668**-255 = 1e44 > 3.4e38``
→ Inf → 前向 NaN(``torch.where`` 只丢结果,不阻止 Inf 被算出来)。
* 朴素 ``γ ** delta`` 对 γ<1 安全(下溢到 0,那里本来就该是 0),
但对 γ>1 的门控不成立;对数形式两端都对。
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
「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
优化目标一律是**减少发射次数**,而不是减少计算量:
* 索引张量进缓存(省 4 次发射);
* 三张一维表复用同一个 ``γ^j``:[H,C]:``cross = γ^j·γ``、
``value = flip(γ^j)``、``chunk = γ^j[-1]·γ``,
把原来的 3 次 ``exp`` 压成 1 次。
"""
causal, expo, relf = _index_tables(chunk_size, device)
g = gamma.to(device=device, dtype=torch.float32)
# γ 可以合法地取到 0(门控被训练成「完全遗忘」),此时 log(0) = -inf,
# 而对角线 delta=0 会让 exp(-inf * 0) = exp(NaN) —— torch.where 只丢结果、
# 不阻止 NaN 被算出来,于是整层 NaN。夹一个正下限后
# exp(logγ·0) = 1、exp(logγ·k) 下溢到 0,正好是参考实现 gamma**delta 的
# 语义(0**0 = 1、0**k = 0),两条路径在 γ=0 处重新等价。
log_gamma = torch.log(g.clamp_min(torch.finfo(torch.float32).tiny))
inner_mask = torch.exp(log_gamma[:, None, None] * expo[None])
inner_mask = torch.where(causal, inner_mask,
torch.zeros((), device=device, dtype=torch.float32))
pow1d = torch.exp(log_gamma[:, None] * relf[None, :]) # [H,C] = γ^j
cross_decay = pow1d * g[:, None] # γ^(j+1)
value_decay = pow1d.flip(-1) # γ^(C-1-j)
chunk_decay = pow1d[:, -1] * g # γ^C
return (inner_mask.to(dtype), cross_decay.to(dtype),
value_decay.to(dtype), chunk_decay.to(dtype))
def chunkwise_batched(q, k, v, gamma, state, chunk_size):
"""Drop-in replacement for ``MultiScaleRetention._chunkwise_retention``.
q, k: [B,H,L,kd] v: [B,H,L,vd] gamma: [H] fp32
state: [B,H,kd,vd] or None -> (output [B,H,L,vd], new_state)
"""
B, H, L, kd = q.shape
vd = v.shape[-1]
if L % chunk_size != 0:
raise ValueError(f"seq_len {L} not a multiple of chunk_size {chunk_size}")
nc = L // chunk_size
C = chunk_size
dev = q.device
inner_mask, cross_decay, value_decay, chunk_decay = _decay_tables(
gamma.to(device=dev, dtype=torch.float32), C, q.dtype, dev)
has_state = state is not None
if state is None:
state = torch.zeros(B, H, kd, vd, device=dev, dtype=q.dtype)
# [B,H,nc,C,*] views — no copy (L is contiguous)
qr = q.view(B, H, nc, C, kd)
kr = k.view(B, H, nc, C, kd)
vr = v.view(B, H, nc, C, vd)
# (1) per-chunk state contributions, one batched matmul over all chunks
U = torch.matmul(kr.transpose(-1, -2),
vr * value_decay[None, :, None, :, None]) # [B,H,nc,kd,vd]
# (2) exclusive scan -> S_prev[i] = state before chunk i
#
# NOTE(2026-10-10): 原来的写法是 nc 次串行 ``s = decay*s + U[:,:,i]``
# 外加一次 ``torch.stack`` —— nc=16 时 33 次内核发射、实测 **591 µs**,
# 而每个参与的张量只有 0.5 MB:几乎全部是发射延迟。
#
# 关键观察是 ``decay = chunk_decay`` 在整个循环里**是常量**(它是 [H] 的
# 每头常数,不随 chunk 变),于是有闭式解
#
# S_prev[i] = d^i · state + Σ_{j<i} d^(i-1-j) · U_j
#
# 这正是对 ``T[h,i,j] = d_h^(i-1-j)``(下三角)的一次批量矩阵乘:
#
# S_all = T[None] @ U.reshape(B,H,nc,kd*vd) [B,H,nc+1,kd*vd]
#
# 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
# §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。
#
# NOTE(2026-10-10): 之前试过对数深度(Hillis-Steele)扫描把启动数从 2·nc
# 降到 ~2·log2(nc),**反而慢 2.8 倍**(3004 vs 1078 ms),内核数还从
# 10116 涨到 11472 —— 因为每轮都要 ``torch.cat`` 物化一个
# [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
# 额外的大张量(``T`` 只有 H·(nc+1)·nc 个元素),所以是另一回事。
# Collected in a list and stacked: writing into a torch.empty() buffer
# with slice assignment makes autograd emit CopySlices and leaves the
# untouched-in-backward buffer as uninitialised memory.
# 同样要夹下限:expo_t 在 (i=0, j=0) 处为 -1,clamp_min(0) 后是 0,
# 若 log_d 是 -inf 则 -inf*0 = NaN 会留在 T 的对角线上。
log_d = (torch.log(gamma.to(device=dev, dtype=torch.float32)
.clamp_min(torch.finfo(torch.float32).tiny))
* float(chunk_size))
idx = torch.arange(nc + 1, device=dev)
expo_t = idx[:, None] - torch.arange(nc, device=dev)[None, :] - 1
valid_t = expo_t >= 0
T = torch.exp(log_d[:, None, None] * expo_t.clamp_min(0).to(torch.float32)[None])
T = torch.where(valid_t[None], T, torch.zeros((), device=dev, dtype=torch.float32))
T = T.to(q.dtype) # [H,nc+1,nc]
U_flat = U.reshape(B, H, nc, kd * vd)
S_all = torch.matmul(T[None], U_flat) # [B,H,nc+1,kd*vd]
if has_state:
t_state = torch.exp(log_d[:, None] * idx.to(torch.float32)[None]).to(q.dtype)
S_all = S_all + t_state[None, :, :, None] * state.reshape(B, H, 1, kd * vd)
S_prev = S_all[:, :, :nc].reshape(B, H, nc, kd, vd)
new_state = S_all[:, :, nc].reshape(B, H, kd, vd)
# (3) inner chunk, one batched matmul
inner = torch.matmul(
torch.matmul(qr, kr.transpose(-1, -2)) * inner_mask[None, :, None, :, :], vr)
# (4) cross chunk, one batched matmul
cross = torch.matmul(qr * cross_decay[None, :, None, :, None], S_prev)
out = (inner + cross).reshape(B, H, L, vd)
return out, new_state