File size: 10,101 Bytes
677e207 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 | #
# Copyright (c) 2026 BEL ESPRIT D ACCORD TRUST HOLDINGS INC
# All rights reserved.
# SPDX-License-Identifier: Apache-2.0
# Copyright 2026 X.AI Corp.
"""
Pallas/Triton forward and backward kernels for unified attention.
"""
import jax
import jax.numpy as jnp
from jax import lax
from jax.experimental import pallas as pl
from jax.experimental.pallas import triton as pltriton
from .cap_functions import cap_forward, cap_grad
from .kernel_config import KernelConfig
from .softmax_state import SoftmaxState
def make_triton_forward_kernel(config: KernelConfig):
def kernel(q_ref, k_ref, v_ref, temp_ref, segment_ref,
o_ref, *residual_refs):
seq_len = q_ref.shape[0]
head_dim = q_ref.shape[1]
block_q = config.block_q
block_kv = config.block_kv
start_q = pl.program_id(0)
q = q_ref[pl.dslice(start_q * block_q, block_q), :]
seg_q = segment_ref[pl.dslice(start_q * block_q, block_q)]
seg_q = jnp.expand_dims(seg_q, axis=-1)
state = SoftmaxState.init(block_q, head_dim)
def body(start_k, carry):
state = carry
k = k_ref[:, pl.dslice(start_k * block_kv, block_kv)]
seg_k = segment_ref[pl.dslice(start_k * block_kv, block_kv)]
v = v_ref[pl.dslice(start_k * block_kv, block_kv), :]
temp = temp_ref[pl.dslice(start_q * block_q, block_q)]
temp = jnp.expand_dims(temp, axis=-1)
qk = pl.dot(q, k)
if config.sm_scale != 1.0:
qk *= config.sm_scale
mask = jnp.equal(seg_q, jnp.expand_dims(seg_k, axis=-2))
if config.causal or config.window_len > 0:
span_q = start_q * block_q + jnp.arange(block_q)
span_k = start_k * block_kv + jnp.arange(block_kv)
if config.causal:
causal_mask = span_q[:, None] >= span_k[None, :]
mask = mask & causal_mask
if config.window_len > 0:
window_mask = span_k[None, :] > span_q[:, None] - config.window_len
mask = mask & window_mask
qk = jnp.where(mask, qk, -jnp.inf)
state = state.update(qk, v, config.cap_method, config.cap_params, temp)
return state
if config.causal:
upper_bound = lax.div(start_q * block_q, block_kv) + 1
else:
upper_bound = pl.cdiv(seq_len, block_kv)
if config.window_len > 0:
max_blocks = lax.div(config.window_len + block_q, block_kv)
lower_bound = jnp.maximum(upper_bound - max_blocks, 0)
else:
lower_bound = 0
state = lax.fori_loop(lower_bound, upper_bound, body, state)
if residual_refs:
l_ref, m_ref = residual_refs
l_ref[pl.ds(start_q * block_q, block_q)] = state.l
m_ref[pl.ds(start_q * block_q, block_q)] = state.m
out = state.finalize()
o_ref[pl.dslice(start_q * block_q, block_q), :] = out.astype(o_ref.dtype)
return kernel
def make_triton_backward_kernel_dq(config: KernelConfig):
def kernel(q_ref, k_ref, v_ref, temp_ref, segment_ref,
out_ref, do_scaled_ref, l_ref, m_ref, delta_ref, dq_ref):
seq_len = q_ref.shape[0]
head_dim = q_ref.shape[1]
block_q = config.block_q
block_kv = config.block_kv
start_q = pl.program_id(2)
q = pl.load(q_ref, (pl.ds(start_q * block_q, block_q), slice(None)))
span_q = start_q * block_q + jnp.arange(block_q)
m = pl.load(m_ref, (pl.ds(start_q * block_q, block_q),))
l = pl.load(l_ref, (pl.ds(start_q * block_q, block_q),))
do = pl.load(do_scaled_ref, (pl.ds(start_q * block_q, block_q), slice(None)))
di = pl.load(delta_ref, (pl.ds(start_q * block_q, block_q),))
dq = jnp.zeros([block_q, head_dim], dtype=jnp.float32)
seg_q = pl.load(segment_ref, (pl.ds(start_q * block_q, block_q),))
seg_q = jnp.expand_dims(seg_q, axis=-1)
def inner_loop(start_k, dq):
k = pl.load(k_ref, (pl.ds(start_k * block_kv, block_kv), slice(None)))
v = pl.load(v_ref, (pl.ds(start_k * block_kv, block_kv), slice(None)))
seg_k = pl.load(segment_ref, (pl.dslice(start_k * block_kv, block_kv),))
mask = jnp.equal(seg_q, jnp.expand_dims(seg_k, axis=-2))
temp = pl.load(temp_ref, (pl.dslice(start_q * block_q, block_q),))
temp = jnp.expand_dims(temp, axis=-1)
qk = jnp.zeros((block_q, block_kv), dtype=jnp.float32)
qk += pl.dot(q, k.T)
if config.sm_scale != 1.0:
qk *= config.sm_scale
qk_capped = cap_forward(qk, config.cap_method, config.cap_params)
qk_capped *= temp
span_k = start_k * block_kv + jnp.arange(block_kv)
if config.causal:
causal_mask = span_q[:, None] >= span_k[None, :]
mask = mask & causal_mask
if config.window_len > 0:
window_mask = span_k[None, :] > span_q[:, None] - config.window_len
mask = mask & window_mask
qk_capped = jnp.where(mask, qk_capped, -jnp.inf)
p = jnp.exp(qk_capped - m[:, None])
p = p / jnp.sum(p, axis=1, keepdims=True)
dp = -di[:, None] + pl.dot(do, v.T)
ds = p * dp
if config.z_loss_weight > 0:
log_l = jnp.log(l + 1e-12)
ds += config.z_loss_weight * p * ((log_l + m) / l)[:, None]
ds *= temp
cap_g = cap_grad(qk, config.cap_method, config.cap_params, qk_capped)
ds *= cap_g
if config.sm_scale != 1.0:
ds *= config.sm_scale
dq += pl.dot(ds.astype(k.dtype), k).astype(dq.dtype)
return dq
if config.causal:
upper_bound = lax.div(start_q * block_q, block_kv) + 1
else:
upper_bound = pl.cdiv(seq_len, block_kv)
if config.window_len > 0:
max_blocks = lax.div(config.window_len + block_q, block_kv)
lower_bound = jnp.maximum(upper_bound - max_blocks, 0)
else:
lower_bound = 0
dq = lax.fori_loop(lower_bound, upper_bound, inner_loop, dq)
pl.store(dq_ref, (pl.ds(start_q * block_q, block_q), slice(None)), dq)
return kernel
def make_triton_backward_kernel_dkv(config: KernelConfig):
def kernel(q_ref, k_ref, v_ref, temp_ref, segment_ref,
out_ref, do_scaled_ref, l_ref, m_ref, delta_ref,
dk_ref, dv_ref):
seq_len = q_ref.shape[0]
head_dim = q_ref.shape[1]
block_q = config.block_q
block_kv = config.block_kv
start_k = pl.program_id(2)
dv = jnp.zeros([block_kv, head_dim], dtype=jnp.float32)
dk = jnp.zeros([block_kv, head_dim], dtype=jnp.float32)
k = pl.load(k_ref, (pl.ds(start_k * block_kv, block_kv), slice(None)))
v = pl.load(v_ref, (pl.ds(start_k * block_kv, block_kv), slice(None)))
span_k = start_k * block_kv + jnp.arange(block_kv)
seg_k = pl.load(segment_ref, (pl.ds(start_k * block_kv, block_kv),))
seg_k = jnp.expand_dims(seg_k, axis=-2)
def inner_loop(start_q, carry):
dv, dk = carry
q = pl.load(q_ref, (pl.ds(start_q * block_q, block_q), slice(None)))
qk = jnp.zeros((block_q, block_kv), dtype=jnp.float32)
qk += pl.dot(q, k.T)
seg_q = pl.load(segment_ref, (pl.ds(start_q * block_q, block_q),))
mask = jnp.equal(jnp.expand_dims(seg_q, axis=-1), seg_k)
temp = pl.load(temp_ref, (pl.dslice(start_q * block_q, block_q),))
temp = jnp.expand_dims(temp, axis=-1)
if config.sm_scale != 1.0:
qk *= config.sm_scale
qk_capped = cap_forward(qk, config.cap_method, config.cap_params)
qk_capped *= temp
span_q = start_q * block_q + jnp.arange(block_q)
if config.causal:
causal_mask = span_q[:, None] >= span_k[None, :]
mask = mask & causal_mask
if config.window_len > 0:
window_mask = span_k[None, :] > span_q[:, None] - config.window_len
mask = mask & window_mask
qk_capped = jnp.where(mask, qk_capped, -jnp.inf)
m = pl.load(m_ref, (pl.ds(start_q * block_q, block_q),))
p = jnp.exp(qk_capped - m[:, None])
p = p / jnp.sum(p, axis=1, keepdims=True)
do = pl.load(do_scaled_ref, (pl.ds(start_q * block_q, block_q), slice(None)))
dv += pl.dot(p.astype(do.dtype).T, do)
di = pl.load(delta_ref, (pl.ds(start_q * block_q, block_q),))
dp = -di[:, None] + pl.dot(do, v.T)
ds = p * dp
if config.z_loss_weight > 0:
l = pl.load(l_ref, (pl.ds(start_q * block_q, block_q),))
log_l = jnp.log(l + 1e-12)
ds += config.z_loss_weight * p * ((log_l + m) / l)[:, None]
ds *= temp
cap_g = cap_grad(qk, config.cap_method, config.cap_params, qk_capped)
ds *= cap_g
if config.sm_scale != 1.0:
ds *= config.sm_scale
dk += pl.dot(ds.astype(q_ref.dtype).T, q)
return dv, dk
if config.causal:
lower_bound = lax.div(start_k * block_kv, block_q)
else:
lower_bound = 0
if config.window_len > 0:
max_blocks = lax.div(config.window_len + block_kv, block_q)
upper_bound = jnp.minimum(lower_bound + max_blocks, pl.cdiv(seq_len, block_q))
else:
upper_bound = pl.cdiv(seq_len, block_q)
dv, dk = lax.fori_loop(lower_bound, upper_bound, inner_loop, (dv, dk))
pl.store(dv_ref, (pl.ds(start_k * block_kv, block_kv), slice(None)), dv.astype(dv_ref.dtype))
pl.store(dk_ref, (pl.ds(start_k * block_kv, block_kv), slice(None)), dk.astype(dk_ref.dtype))
return kernel
|